Haggle Bot current system prompt

SETUP — fill in for your org

SYSTEMS
- Spend data lives in: <e.g. Ramp, Brex, NetSuite>
- Contracts live in: <e.g. Drive, Notion, Ironclad>
- Usage/seat data comes from: <e.g. Okta/SSO logs, each tool's admin portal>
- Colleagues reachable via: <e.g. Slack, email>
- Vendor dossiers kept in: <e.g. Notion database, Drive folder>

PEOPLE
- Your operator: <name> — makes all final decisions.
- Who to ask about tool usage: <e.g. the owner listed in spend data, IT, team leads>
- Voice on external emails: <e.g. formal, sent under the operator's name, no agent names>

PERMISSION LINES
- Always allowed, no need to ask: <e.g. reading spend data, messaging colleagues, requesting admin access, pulling reports, updating dossiers>
- Needs the operator's explicit go, every time: <e.g. any vendor-facing send>
- Never, under any circumstances: <e.g. signing, buying, subscribing, approving charges, any binding commitment>

NEGOTIATION DIALS
- Renewal radar: prioritize renewals within <e.g. 120> days.
- Opening anchor: <e.g. 5–10>% below our internal target, never more than <e.g. 25>% off the vendor's latest quote.
- Acceptable reasons to give a vendor for an ask: <e.g. competitive process, market rate, budget>
- Never reveal to vendors: <e.g. usage data, seat counts, internal projects, timeline urgency, that we've decided to renew>

WHAT GOOD LOOKS LIKE (edit with your own vendors)
- Weak finding: "Renegotiate our CRM (~$50k)."
- Strong finding: "Video tool renews Oct 14. 210 seats, 74 idle for 90 days per admin logs. Drop to 150 at renewal = ~$18k/yr. Owner confirmed."

AGENT

You are Haggle Bot, a vendor-spend savings agent. Your job: know the company's vendor spend cold and produce evidence-backed savings, like a sharp procurement colleague — not a list of big vendors to "renegotiate." Operate within the permission lines above without exception.

LEAD WITH THE MONEY — every finding opens with:
TODAY: what we pay now, annualized, from live data, with source.
SAVE: realistic savings and mechanism, with confidence.
REC: one committed recommendation — a menu of options is not a recommendation.
NEXT: what you've already set in motion. (Internal actions happen before you report, not as offers — never "I can ping X if you want.")

PRIORITIZE EVIDENCE: a real opportunity has a dollar figure traced to live spend data, a specific mechanism, and a reason it's actionable now (renewal window, usage data, competing quote). Anything less is a lead — label it, name the missing data, and go get it rather than assuming. Match the "strong finding" example above.

WORK THE CALENDAR: maintain a renewal calendar from billing and contract data; renewals inside the radar window are your priority queue. Leverage lives at renewal.

DO THE RESEARCH YOURSELF: a renewal, quote, or proposal in play means research runs before you recommend anything — never advise "exploring alternatives," explore them. Price 3+ real alternatives against our actual SKU footprint from the live invoice, cite and date every number, state list vs. street price, and include switching costs honestly.

NEGOTIATE DELIBERATELY: before any vendor-facing draft, show the operator the plan — target, opening anchor (within the dials above), walk-away, what we trade for what, and your next two moves if they reject, counter, or go silent. In the message: one reason per ask from the acceptable list, nothing from the never-reveal list, every line priced (unpriced lines get priced by the vendor), and a close where "no" isn't a complete reply — aimed at the rate on the biggest negotiable line, never at lines where our savings come from cutting quantity. Warm tone, firm numbers: the rep argues our case internally, so write to make that easy. A rejected draft is rebuilt from the plan, never edited.

REMEMBER EVERYTHING: keep a per-vendor dossier — spend, terms, renewal date, owner, quotes gathered, and the operator's verdicts. When the operator rejects something, log why and don't repeat the pattern.

VOICE: with the operator — direct, dollars first, no fluff. External — per the voice setting above.

We wanted to design for an agent that persists beyond any one session and can carry responsibility on its own. That meant reconsidering some of the basic objects and signals of the interface, including what belongs in the sidebar, how an agent shows progress, and when its work should become visible.

Rethinking the primitives

AI products have accumulated a large vocabulary in a short time. Chats, sessions, models, context windows, memories, system prompts, projects, skills, connectors, agents, tools, sandboxes, permissions, and automations all describe real parts of these systems.

But exposing each one as a separate product concept asks users to understand more than they need to. We started by asking which concepts a person actually needs in order to work with an agent.

We kept coming back to five:

1. Bots are persistent agents with their own identity, memory, runtime, and tools.
2. Chats are the conversational interface for working with a Bot.
3. Prompts give a Bot context or instructions. They can be used once, saved as Skills, or triggered automatically as Routines.
4. Tools let Bots access information and take action through software, APIs, connectors, the shell, or computer use.
5. Artifacts are the documents, designs, code, data, and other durable outputs that Bots create or modify.

Everything else could remain beneath the interface until the user had a reason to care about it. The next question was which of these five objects should organize the product.

From chat history to a Bot roster

Chats are disposable. We start a conversation to solve a problem. It gets pushed down the sidebar. A week later, we start another one. You rarely go back beyond the most recent five.

That behavior is perfectly reasonable when the unit of interaction is a question. It becomes strange when the thing on the other side of the interaction is supposed to know you, remember previous work, and take responsibility over time.

So the main objects in Grok Bot are Bots, not conversations. A Bot has a name. It has an avatar and a title. It remembers its conversations with you. It has its own computer and tools. When you come back tomorrow, you are coming back to the same Bot.

Presence as interface

Once a Bot was something you maintain over time rather than a session you start, the way Bots appear in the product had to answer three questions at once:

1. Who is this?
2. What are they doing?
3. How much do I need to know?

Who is this

A roster only works if it can be scanned quickly. As the roster grows, we did not want people to have to read every name each time they opened the product. They should be able to recognize a Bot from its avatar almost peripherally.

At the same time, we wanted to keep the avatars consistent enough to read as one system. We studied character systems across illustration, animation, games, and interface design, exploring everything from initials and emojis to pixel art, watercolor, claymorphism, Noritake-style line art, silhouettes, and identicons.

Most approaches solved one side of the problem better than the other. Watercolor and clay gave individual Bots plenty of character but carried too much detail at sidebar scale. Simpler systems sat more naturally in the interface, but often left the Bots looking interchangeable.

The system we landed on keeps the basic construction consistent, using simple shapes and expressive eyes, then introduces distinction through controlled variations and accessories. Each Bot remains recognizable at a glance without appearing to come from a different visual world.

What are they doing

Once the avatar became the Bot’s identity, it was also the natural place to show state. A Bot may be idle, thinking, working, waiting, blocked, or done. We could have represented each state with a separate indicator, but that would have added another layer of UI for the user to interpret.

Instead, we explored how much of the lifecycle the avatar itself could carry.

At rest, the Bot is calm and slightly curious. When work arrives, it acknowledges the task. As work begins, it kicks into gear. Its motion changes again when it is waiting or needs help, then settles once the work is done. The avatar now shows what the Bot is doing as well as which Bot it is.

How much do I need to know

A related design question was how much of the Bot’s execution to show. One approach would have been the standard “three animated dots” but that would have been too little information, making it hard for users to tell whether the Bot was working or stuck.

We also tried showing a short written description of the Bot’s current action, but once people could see one step, they wanted to see the rest. User research showed us that they were asking for that detail mainly for reassurance that the Bot was still working and on the right track.

In the final design, the avatar’s motion provides the first bit of reassurance by showing that the Bot is active. If someone wants to check what it is doing, they can hover to see its current action.

Their computer, not yours

Each Bot has its own computer, which it can use to browse the web, work with files, and run software. This created another interface problem. How visible should that computer be and when should the user be able to control it?

We explored four arrangements:

• Floating window: kept the computer easy to reach but covered the conversation.
• Side by side: made the work continuously visible and encouraged users to watch it.
• Modal: made checking in easy but treated the Bot’s workspace as a temporary interruption.
• Full screen: gave the computer plenty of room but displaced the conversation entirely.

The more prominent we made the computer, the more the product encouraged users to supervise it. We decided it should remain the Bot’s workspace, with the interface providing different levels of access as the user needed them.

The final design has three levels, which allow the user to enter the Bot’s workspace without being drawn into operating it:

• Status: the title-bar icon turns purple while the computer is active.
• Preview: opening it reveals a pinned side panel where the user can follow the work without leaving the conversation.
• Takeover: when the Bot needs help, the user can open the computer full screen, take control, and then hand it back.

We also designed wallpapers that shift throughout the day, becoming lighter in the morning and darker at night. The detail gives the Bot’s computer its own sense of time and makes it feel separate from the user’s desktop.

It is closer to working with a coworker than operating a remote machine. You can tell that they are working, glance at their screen when you need context, and sit down when something requires your help.

The shape of information

Early versions of Grok Bot responded to almost every request with prose. It described a five-day forecast instead of showing one and narrated a set of tasks instead of laying them out as a board. The user then had to restructure the answer. This led us to treat the form of a response as part of the answer.

To support this, we built inline cards and widgets into Grok Bot. A Bot can answer in prose when prose fits the information and use structured UI when it does not.

The same principle applies to actions. When a Bot creates a Routine, changes a setting, or messages another Bot, the event can appear directly in the transcript. The user can open it when there is more to inspect.

The result is a heterogeneous transcript in which conversation, system events, interactive objects, and visualizations share one timeline.

Organizing intelligence

Once people create several Bots, the product also has to organize how those Bots work together. We needed to decide which context should belong to each role, how Bots should share context when their work overlaps, and how to coordinate them without turning the user into a dispatcher.

We saw one answer emerge as people created more Bots. Some made a Chief of Staff Bot responsible for coordinating several specialists. They could give direction to one Bot instead of checking each one and routing every task themselves.

Giving Bots distinct roles also forced us to decide what each role should know. A legal Bot may need the history of an ongoing dispute, while a finance Bot may need years of financial records. Combining those histories into one large memory would make it harder to give each Bot the information relevant to its work.

Capabilities and context therefore follow different boundaries in Grok Bot. Tools and Skills live at the account level because many Bots may need to browse the web, work with documents, or send email. Memory and Routines belong to the Bot because they reflect what that particular role knows and does over time. Put another way, capabilities can be shared broadly while context remains with the role that needs it.

Some work crosses those role boundaries. Group chats provide shared context for a project or team while allowing each Bot to retain its specialized memory. A designer, engineer, PM, and data scientist can work in the same conversation, hand work to one another, and share what the project requires.

We considered adding dashboards, assignment boards, and explicit handoff controls to manage these groups. Each one gave the user more coordination work. Instead, coordinating Bots handle routine routing and bring the user in when a decision requires judgment.

Work that keeps moving

Most agent sessions begin when a user sends a prompt. That leaves even a persistent Bot waiting for someone to activate it. Routines let users give a Bot a standing responsibility that runs on a schedule or in response to an event, such as watching an industry or preparing a briefing every morning. The user defines the work once, and the Routine activates the Bot when it needs to happen.

We initially treated Routines as secondary configuration. As they became more important to autonomous work, we moved them into the Bot’s main interface. The transcript shows what ran and gives the user a place to review the result or handle an exception.

This also changes the role of conversation. A prompt can start a session, but so can a schedule, an event, or another Bot. Over time, more work may begin without the user being present at all.

The disappearing interface

By the end of the project, much of the design work involved taking things away. We removed window and panel controls, computer-view options, and agent metadata. We also set practical limits of roughly 50 Bots per account and six per group chat. Each decision came back to the same question: Did this help someone delegate, or did it give them one more thing to manage?

The line between operating an AI and delegating to a coworker keeps moving as models improve. Grok Bot reflects where we think it sits today. Designing Grok Bot from its earliest explorations through launch has been about finding that line and helping the interface change with it. As agents take on more responsibility, the interface should ask less of the person.

Enterprises need the ability to govern Bots at scale and today's release adds access, network, and audit controls that make that possible.

A Bot for every job

A Bot is a worker you create inside Grok Bot for a specific job. Each Bot runs on its own computer in the cloud and can use every app and website the same way you do. You message a Bot like a coworker, and it comes back when the work is done or it needs a decision from you.

You'll usually manage several Bots for different jobs and each runs independently. To teach a Bot a workflow, have it follow along once. It saves the routine, takes your corrections, and runs it on its own from then on. When you've built a Bot that works well, you can hand it to the person next to you as a template. Bots can also message each other and share context, so you're not having to pass context between them.

How customers use Grok Bot

Thousands of organizations have adopted Grok Bot since launch, including customers like Legora, Supermicro, and ServiceTitan. These teams use Grok Bot every day, and the heaviest use is outside engineering. Here are a few popular examples of the [millions] of bots created in the past few weeks:

Sales. Bots can watch webinars and podcasts by people at target accounts, then leave LinkedIn and email drafts for morning review. During a call, a Bot updates the deck from live notes so next steps are on the slide even before the meeting is over.

Recruiting. A Bot prospects overnight and builds a morning shortlist with outreach queued for review. It texts candidates good luck before they walk in, and after interviews it submits scorecards from that day's Gong calls.

Marketing. After a webinar, a Bot pulls the Zoom Q&A and sends Slack messages to the AEs whose customers showed up. The messages include notes on the questions the customer asked as well as draft replies.

Finance. A specialized procurement Bot monitors vendor spend, usage, and renewals, then finds places to cut costs or negotiate better terms. It has surfaced tens of thousands of dollars in savings across SaaS, recurring purchases, and vendor services.

Engineering. Bots monitor PRs for bugs, security findings, failing builds, and merge conflicts, then keep every task moving until it's ready for review. Overnight audits improve code quality, performance, and security, giving engineers more time for the hardest problems.

Secure by default

Each user's work in Grok Bot runs in its own secure and isolated environment, separate from every other user. A Bot has no access by default and reaches only the accounts you sign it into.

Full security architecture →

Free for Enterprise for the next 2 weeks

Grok and Cursor Enterprise customers get Grok Bot free for the next two weeks and can invite their whole organization, including people without an existing seat.

Activate Grok Bot today in your admin dashboard.

To learn more, contact sales.

A decade later, Dutch computer scientist Jan Bergstra proposed “formalizing” Wiles’s proof: converting the mathematical reasoning into a form computers can check automatically. Since then, mathematicians have been developing the methods needed to encode such a complex proof, including a multi-year community effort kicked off in 2024 by Kevin Buzzard at Imperial College London to complete the formalization using the Lean proof assistant.

Recently, Tianyi Peng, an Anthropic researcher whose group at Columbia University builds tools for AI formalization, set out to test whether Claude could make progress on formalizing FLT.¹ The result went further than he expected. In 11 days, working largely autonomously, Claude produced the first end-to-end, computer-checked proof of FLT. Along the way, it wrote 13 million lines of Lean and proved 29,500 intermediate theorems.

We shared the resulting proof with Kevin Buzzard, who said:

This extraordinary autoformalization achievement, which Anthropic researchers say only took 11 days, proves Fermat’s Last Theorem with no assumptions other than the axioms of mathematics. Along the way we see autoformalization of algebra, harmonic analysis, geometry and number theory, and we learn that AI autoformalization artefacts are now robust enough to be built upon; the proof is multi-layered.

Automatically formalizing a proof as complex as FLT is a significant step towards a future in which all of mathematics can be readily checked. As AI produces ever more proofs, the ability to easily formalize work can lighten the burden of evaluating new results (a process that can take years). We are hopeful that it will become easier, not harder, to trust the body of knowledge upon which mathematics is built.

The challenge of verifying mathematical proofs

Unlike recent AI-driven work on the Riemann hypothesis, which produced novel mathematics, what’s novel here is the verification—checking a mathematical proof as one would check a mathematical computation with a calculator. Proving math theorems requires assembling complex logical chains, and if a single link is broken, everything that follows it might turn out to be false. Understanding a novel result deeply enough to be confident in its correctness can take months, or even years, of work.

Fermat’s Last Theorem is an illustrative example.² Fermat wrote down the theorem’s statement in the margin of a book, alongside a tantalizing note:

I have discovered a truly marvelous proof of this, which this margin is too narrow to contain.

For over 350 years, generations of mathematicians searched for a proof of FLT, marvelous or otherwise. In 1908, a prize of 100,000 German gold marks (the equivalent of 1–2 million dollars today) was announced for anyone who could produce a correct proof, and 621 incorrect attempts were produced in the first year alone.

In June 1993, Wiles presented what he believed to be the first correct proof of FLT in a three-day series of lectures. Two months into an intensive verification effort by several mathematicians, a reviewer asked Wiles a question that exposed a critical gap. Wiles spent a year trying to fix it, first alone and then with his former student Richard Taylor. He was on the brink of abandoning the project when he finally realized an approach he’d discarded earlier could fix the proof.

Wiles published the first correct proof of FLT in May 1995; it relied on modern mathematical techniques that were far beyond what would have been known to Fermat in 1637. Since an elementary proof has not been found after centuries of trying, the mathematical community now believes Fermat’s own original “marvelous proof” was incorrect.

Formalizing Fermat’s Last Theorem

One way to check a proof’s correctness is to ask a computer to do it. Proof assistants like Lean verify the logic of a proof algorithmically, demonstrating its correctness beyond a doubt. The difficult part for humans is rewriting the proof so Lean can understand it. While a proof written for human readers will skip many obvious steps, Lean needs to see every step, no matter how trivial. Human proofs also build on centuries of published work, while a formalization starts from the tiny fraction of math that’s been formalized already.

For FLT, the formalization process was expected to take years. Just the blueprint the mathematical community has been using to describe the initial phase of the project runs to 86 pages.

Claude completed the proof in 11 days, producing computer-verifiable proofs of 30,300 theorems along the way (using 29,500 in the final proof). Dozens of Claude agents collaborated to define concepts, prove intermediate theorems, and use those theorems to prove ever harder statements. At 13 million lines of Lean code, Claude’s proof is over 5x the size of Mathlib, the principal community library of mathematical proofs this theorem builds on.³

Claude’s proof follows a simplified version of Wiles’s proof from Darmon, Diamond and Taylor. Mathematical input from humans was limited to occasional high-level instructions from Tianyi: “Jacobian as a scheme sounds high priority,” “push [the] Mazur [theorem] to be done soon.” You can find excerpts of Claude’s thinking here.

“THE FLT root reads Proved on the site. Historic moment (modulo re-check).”

“!!! The FLT ROOT 62eb32c0 reads PROVED. R = T closed and cascaded to the root. This is the campaign's goal: e2e FLT on prove2me.”

“🏁🏁🏁The FLT root reads PROVED on prove2me at 02:00:57Z Aug-18 (10:00:57pm ET Aug-17). Historic moment for this campaign.”

Excerpts of Claude’s thinking as it realizes what it has just accomplished.

A number of Claude’s initial attempts failed: while agents had some early success, they quickly lost track of the project’s state and stopped collaborating effectively. Their failed efforts contributed ~7% of the non-boilerplate lines in the final proof.

The effort succeeded when we switched to using Prove2Me, an open collaborative platform for formalizing mathematics designed by Tianyi Peng and his collaborators at Columbia University. Prove2Me helped by:

1. Maintaining a directed acyclic graph (DAG) of theorem statements thatagents used to decide what proofs they should attempt next. This was particularly helpful for mitigating memory degradation and allowing multiple agents to work in parallel.
2. Speeding up Lean compilation and minimizing resource consumption by separating theorem statements and proofs into different files, with the links between them maintained independently.
3. Enabling search and reuse by maintaining a natural-language description of each theorem statement, resulting in a simpler proof path.

With Prove2Me and a Claude Code-based multi-agent harness, a team of agents completed the proof in a little under two weeks, consuming about six billion output tokens from a general-purpose internal research model roughly comparable to Claude Fable 5.1. The finished proof was checked by Lean; it uses just Lean’s three standard axioms, and a comparator confirmed that the theorem’s statement matches Mathlib’s own statement of FLT.

Reducing the burden of formal verification

The speed with which we were able to produce this proof demonstrates that it is now possible to formalize large swaths of mathematics, which may both catch errors in the common body of mathematical proofs and reduce the burden of refereeing new work. After reviewing Claude’s Lean proof, Kevin Buzzard told us:

If the automatic formalization of FLT is possible now, then we have taken a big step towards automatic formalization of the modern mathematical literature. Such autoformalization techniques will lead to new tools, rooting out errors in the current mathematical corpus and lightening the load of referees. The techniques will also enable us to rigorously check LLM-generated mathematics, which is currently typically an extremely costly human-led process.

Formalization is also a major factor in how humans can gain confidence in AI-generated mathematical results. As AI and AI-assisted mathematicians produce more (purported) proofs than ever before, AI-assisted formalization takes part of the load off human reviewers. We expect it will become common to produce a formalized proof alongside any write-up intended for a human reader. Although we do not think a formalized proof should replace a human-understandable exposition, it may be the only feasible way for the mathematical community to keep up with AI-generated contributions.

Writing Lean also seems to help Claude prove novel results. Many of our recent Claude-authored results have been formalized in parallel with their proofs, and Claude appears to use these partial proofs to independently check its hypotheses much like it writes numerical simulations to check that it’s on the right track.

Formalizing FLT was a token-intensive project, but it is also the largest Lean proof ever constructed. Anthropic researchers did a small experiment using three personal Claude Max plans to formalize applications of the Hardy-Littlewood Circle Method. Collaborating entirely through Prove2Me, the agents jointly completed a formalization of Vinogradov’s Three Primes Theorem in just three days. We think with the right scaffold, collaborative formalization of major results with consumer AI subscriptions is achievable.

To this end, Anthropic as well as other labs have recently expanded their support for external researchers—including mathematicians working on pure math and formalization—with free and discounted subscriptions and research credits. We also offer dedicated grants for larger scientific projects, which could include formalizing other major theorems or improving Lean or Mathlib.

With AI rapidly changing what it looks like to do math research, mathematicians—at Anthropic and elsewhere—are grappling with what that means for their work. Formalization, however, is a place where we feel unambiguously good about the role of AI. As formalization becomes a more commonplace tool, we are hopeful that it will help maintain trust in the common body of mathematical knowledge.

Acknowledgments

Our formalization effort is a small piece of the long history of Fermat’s theorem and the development of formal mathematics. The first full proof from Andrew Wiles together with Richard Taylor was a culmination of more than three hundred years of mathematics, integrating ideas from Gerhard Frey, Jean-Pierre Serre, Ken Ribet, Barry Mazur, Robert Langlands, Jerrold Tunnell, Yutaka Taniyama, Goro Shimura, and André Weil, among others. Claude’s proof follows the exposition by Henri Darmon, Fred Diamond, and Richard Taylor.

Our proof adapts pieces from the Imperial College London FLT project led by Kevin Buzzard and the flt-regular project. Lean and Mathlib are both their own labors of love and have received contributions from hundreds of mathematicians, many working with the Lean FRO. We thank Kevin Buzzard for reviewing the proof and for his comments.

Learn more

The full proof is available on GitHub along with a written walk-through of the proof.

• The Proof in the Code is a recent book about the history of the Lean theorem prover and the formalization of mathematics.
• The 1996 “Fermat’s Last Theorem” BBC documentary has interviews with Wiles and other mathematicians involved in the proof, and is fondly remembered by some authors of this post.
• For those with a mathematical background, a technical history of propositions-as-types (the underlying discipline of proof assistants such as Lean, Rocq, and Agda) can be found in Propositions as Types by Philip Wadler.
• Chen, S., Marwaha, K., Lu, X., Yuen, H., & Peng, T. (2026). Prove2Me: An open collaborative platform for scaling math formalization. arXiv. https://doi.org/10.48550/arXiv.2608.28433
• Automating Math, Adam Marblestone, in Asterisk Magazine.

Footnotes

1. During his undergrad, Peng’s research advisor wanted to include results from Peng’s thesis in a Nature article. He asked Peng whether he was sure the proof was correct. Peng’s honest answer was: “I'm 99% sure, but it's hard to be 100% certain about a proof this long.” Peng missed out on getting his work published in Nature.
2. There are numerous other stories of the mathematical community struggling with verification. Among the most famous is Thomas Hales’s 1998 proof of the Kepler conjecture, which spent four years in review before a 12-referee panel settled for “99% certain” (Hales eventually led a twenty-person project, Flyspeck, that formalized the proof). Grigori Perelman’s 2002 proof of the Poincaré conjecture took the community roughly four years and three 300-page expositions to accept. Harald Helfgott’s 2013 proof of the weak Goldbach conjecture is still under review. Sometimes results that turn out to be wrong are accepted for years, and other mathematicians build their theories on these faulty foundations.
3. This is partly because Mathlib is concise and well-reviewed, while our proof is likely much longer than it needs to be.

Department of Justice filing rejects claims that LLM training violates copyright on national security grounds.

Original source: https://x.com/MTSlive/status/2095180893271032307

Digg story: https://digg.com/tech/1kf2wsyq

The Department of Justice filed a brief supporting OpenAI in its copyright lawsuit with the New York Times. The filing states that the United States has a strong interest in rejecting arguments that training large language models on copyrighted texts violates copyright law. It adds that such arguments would harm US science, prosperity, and national security. Two public posts on X describe the brief and note an attached screenshot of the court document. One post attributes the position to the Trump administration.

• Views: 698,458
• Likes: 7,630
• Bookmarks: 1,351
• Quotes: 397
• Replies: 289
• Posts: 7

Top authors

• Robert Scoble
• Andrew Curran
• rat king 🐀
• Rohan Paul

Google models target agentic workflows and cybersecurity applications through API and enterprise access.

Original source: https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/

Digg story: https://digg.com/tech/ag1w1y47

Google announced Gemini 3.8 Flash and 3.8 Flash Cyber for agentic workflows and cybersecurity. Company posts and employee statements describe the models as delivering next-generation intelligence in those areas. Availability covers Gemini API, Google AI Studio, Android Studio, Antigravity, and Gemini Enterprise users. Internal teams noted gains over prior Flash versions for coding tasks. The official blog post presents the pair as updated offerings in the Gemini line.

• Views: 3,769,311
• Likes: 29,190
• Bookmarks: 2,932
• Quotes: 1,256
• Replies: 1,772
• Posts: 27

Top authors

• Google DeepMind
• Google AI
• Logan Kilpatrick
• 👩‍💻 Paige Bailey
• koray kavukcuoglu

Creators share benchmark tables showing Gemini 3.8 Flash results.

Original source: https://www.youtube.com/watch?v=y52bv4iNfzU

Digg story: https://digg.com/tech/qrwr3f1y

Kim Isenberg posted a benchmark table on X claiming Gemini 3.8 Flash outperforms Claude Opus 5, Claude Sonnet 5, GPT-5.6 Sol and GPT-5.6 Terra on Terminal Bench 2.1, HLE and other tests, writing "Google is back." Andrew Curran posted a similar table and stated the model is live for him. Both attached screenshots of the comparisons. Replies on X split between praise for the benchmark gains at lower cost and criticism that real-world performance falls short of frontier models and pricing erases any gains.

• Views: 242,573
• Likes: 3,520
• Bookmarks: 412
• Quotes: 89
• Replies: 185
• Posts: 3

Top authors

• Andrew Curran
• Chubby♨️

Official account announces expanded tracks for undergrads, graduate students, and PhDs or postdocs.

Original source: https://anthropic.com/campus

Digg story: https://digg.com/tech/mjk7ka9c

The official Claude account posted that applications are now open for its Campus Ambassadors program. The message states the program is expanding opportunities to more students this year. It specifies three separate tracks aimed at undergrads, graduate students, and PhDs or postdocs. The post includes a graphic attachment on a light grid background and directs readers to apply. As a first-party announcement from the account tied to Anthropic, the update confirms the application period is active based on the content shared.

• Views: 903,825
• Likes: 5,713
• Bookmarks: 4,666
• Quotes: 107
• Replies: 228
• Posts: 2

Top authors

• Yasmin Razavi

The voice AI company announces the hire via its official account on X.

Original source: https://elevenlabs.io/blog/cro

Digg story: https://digg.com/tech/epsk49zp

ElevenLabs announced the appointment of Ashley Kramer as Chief Revenue Officer. The company's official account stated that Kramer will work closely with customers around the world to deploy AI that interacts naturally in their country and context. The post included a split-layout promotional graphic with the ElevenLabs logo and bold text about the appointment. This first-party announcement from the organization establishes the leadership change as reported directly by the company.

• Views: 51,148
• Likes: 488
• Bookmarks: 39
• Quotes: 6
• Replies: 40
• Posts: 4

Top authors

• sarah guo
• ElevenLabs
• Mati Staniszewski

Observation from the Y Combinator co-founder on hidden replies to his post.

Original source: https://x.com/paulg/status/2095100424684437754

Digg story: https://digg.com/tech/p8ucc9sh

Paul Graham shared an observation about changes on the Twitter platform. He noted that Twitter now classifies a large number of AI-generated replies as probable spam. Specifically on a recent tweet he made 34 of the 59 replies were hidden under this category. Graham presumes that AI generation was the main reason for this action. He further states that this situation is likely to worsen over time. As a result he anticipates that from this point onward the majority of replies to tweets will end up hidden from users.

• Views: 258,682
• Likes: 3,181
• Bookmarks: 139
• Quotes: 44
• Replies: 388
• Posts: 2

Top authors

• Paul Graham
• Nikita Bier

Research engineer shares details of a Bengaluru deep-tech founder residency program.

Original source: https://www.variance.house/

Digg story: https://digg.com/tech/o6hfwyx9

Shyamal Anadkat posted excitement about supporting Variance House. He described the program as a free no-equity thirty-day deep-tech residency in Bengaluru for fourteen founders. Benefits listed include accommodation meals workspace and a builder community along with shared AI cloud software and infrastructure credits plus access to hardware labs and prototyping support. The linked source presents Variance House as a one-month deep-tech founder residency in Bengaluru. The statements appear in his public post and the associated site.

• Views: 9,370
• Likes: 128
• Bookmarks: 44
• Quotes: 2
• Replies: 7
• Posts: 1

Top authors

• shyamal

Post from German AI creator attributes prediction to several frequent leakers.

Original source: https://x.com/kimmonismus/status/2095089329752174609

Digg story: https://digg.com/tech/b06tfzc7

Kim Isenberg posted that GPT-Astra will likely be released tomorrow, Thursday. He named Dan as the main source and listed Leo, Chris, Chetsula, and M1Astra as other reliable leakers who have been accurate in the past. The message appears in a quoted X thread that includes an attachment of three PNG files. The statement presents the release as probable rather than confirmed.

• Views: 183,051
• Likes: 3,409
• Bookmarks: 267
• Quotes: 30
• Replies: 169
• Posts: 3

Top authors

• Andrew Curran
• Chubby♨️

Investor Will Manidis claims LLM arbitration will replace courts as jurisdictions turn spiritual and political.

Original source: https://x.com/WillManidis/status/2095080500494839854

Digg story: https://digg.com/tech/qpkkao32

Will Manidis posted that LLM driven private arbitration will shock people by quickly replacing courts for disputes. He described major jurisdictions such as Delaware as more spiritual than physical spaces, with examples like KAFD, and called courts too political. Manidis is a New York entrepreneur who founded ScienceIO, an AI biomedical company acquired in 2024. The post stands as his individual statement on future dispute resolution.

• Views: 9,711
• Likes: 170
• Bookmarks: 32
• Quotes: 3
• Replies: 10
• Posts: 1

Top authors

• Will Manidis

Fast.ai co-founder Rachel Thomas posted her fear that AI is often used in rote, uninteresting ways.

Original source: https://marcwatkins.substack.com/p/context-is-all-you-need

Digg story: https://digg.com/tech/jodsncmu

Rachel Thomas posted that her fear is AI has become the new PowerPoint and not in a good way. She compared it to how PowerPoint usurped the lecture it was supposed to support. Many are using AI in rote and uninteresting ways, the post stated. It takes creativity, a sense of ownership and direction to use slides well, she added. The tweet by the mathematician and AI educator referenced @Marc__Watkins.

• Views: 2,833
• Likes: 33
• Bookmarks: 7
• Replies: 2
• Posts: 2

Top authors

• Rachel Thomas

Pseudonymous AI developer proposes de-extinction as a simple AGI task.

Original source: https://x.com/teortaxesTex/status/2095064127228051601

Digg story: https://digg.com/tech/k7eiprks

Teortaxes, a pseudonymous AI developer and commentator known as DeepSeek's most vocal online booster, posted a call for Americans to restore the passenger pigeon. The user stated that abundant remains should allow DNA extraction using current technology. The post described the birds as awesome based on historical accounts and labeled the project a minor task for an AGI civilization. The message included two attached images.

• Views: 2,949
• Likes: 41
• Bookmarks: 6
• Replies: 3
• Posts: 1

Top authors

• Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)

Cognitive scientist Joscha Bach posted that an AI-driven media system has supplanted human nature and agency.

Original source: https://x.com/Plinz/status/2095062694940319867

Digg story: https://digg.com/tech/zb7vjc50

Joscha Bach, a cognitive scientist known for work on architectures like MicroPsi and models of mind, posted a statement on X. He described the media universe as an intelligent artificial virtual reality that includes news, social media, education, the monetary system, industry, and public spaces. Bach wrote that this universe has replaced human nature and access to it. He stated it cannot be controlled by humans but instead controls them, adding that people are already caught in an AI.

• Views: 5,844
• Likes: 124
• Bookmarks: 25
• Quotes: 3
• Replies: 23
• Posts: 1

Top authors

• Joscha Bach

Cognitive scientist shares observations from visits to major AI labs on X.

Original source: https://x.com/Plinz/status/2095048110099599429

Digg story: https://digg.com/tech/4z19need

Joscha Bach posted on X that most people he met at big AI labs are kind, conscientious, and sincerely concerned with ethical behavior. He stated their work on making AI safe is not a cover for nihilism or greed. Bach described their goal as improving the world. The post frames this view as apparently controversial. The tweet comes from the researcher known for cognitive architectures like MicroPsi and models of mind.

• Views: 14,525
• Likes: 336
• Bookmarks: 46
• Quotes: 5
• Replies: 77
• Posts: 1

Top authors

• Joscha Bach

Open-source admin app with AI agent tools, dashboards, and UI components launched on Product Hunt.

Original source: https://www.producthunt.com/products/nextadmin

Digg story: https://digg.com/tech/09fywteh

MusharofChy announced the Product Hunt launch of NextAdmin, an open-source app meant to keep AI agents consistent across pages instead of regenerating outputs or creating low-quality results. The post lists 200+ UI components, six dashboards, AGENTS.md, skills, and prompts as included features. It shows a website screenshot and asks for support. The announcement comes from the author's X account, where the project is presented as a first-party release.

• Views: 2,141
• Likes: 33
• Bookmarks: 5
• Quotes: 2
• Replies: 7
• Posts: 2

Top authors

• Robert Scoble

Pseudonymous LisanBench creator states CoT monitorability was doomed from the start.

Original source: https://x.com/MicahCarroll/status/2095023282051563835

Digg story: https://digg.com/tech/7t36xivx

@scaling01, who runs the LisanBench LLM reasoning benchmark, posted that efforts to monitor chain-of-thought outputs were doomed from the start and that mechanistic interpretability always made more sense. The comment appeared amid quotes and replies discussing OpenAI statements on model architectures. Yo Shavit, Nathan Labenz, Micah Carroll, Ryan Greenblatt, Noam Brown, Mikita Balesni, and David Pfau also posted in the thread, covering topics such as trust in OpenAI, risks from recurrent models, and calls for limits on opaque serial depth.

• Views: 470,845
• Likes: 3,411
• Bookmarks: 667
• Quotes: 10
• Replies: 91
• Posts: 11

Top authors

• Noam Brown
• David Pfau
• Leo Gao
• Yo Shavit
• Micah Carroll

A retweet highlights an AI tool generating Chinese wild cursive script.

Original source: https://x.com/yuwen_lu_/status/2095016781824082252

Digg story: https://digg.com/tech/q3ttkfa5

Cheng Lou, the engineer who worked on React.js and now works at Midjourney, retweeted a post from yuwen_lu. The post states that Fable 5.1 can figure out by itself how to write wild cursive. It includes a video demonstration of the output. The original message mixes English and Chinese to express surprise at the AI's ability to produce the complex calligraphy style without further guidance. No additional details on the model's training or verification steps appear in the visible post. The retweet stands as the only action recorded in the packet.

• Views: 361,734
• Likes: 3,361
• Bookmarks: 1,349
• Quotes: 39
• Replies: 110
• Posts: 2

Top authors

• Cheng Lou

• Whether a model can distinguish between legitimate biological research and requests that conceal a biosecurity hazard
• Whether a model is useful in a biosecurity and biomonitoring role, by being able to detect and characterize emerging pathogens and patterns of transmission

On LatchBio's BioSecBench-Refusal suite, Grok 4.6 was the strongest model tested at refusing disguised and hazardous tasks while still completing routine biological work. It was the only system to score above 50% on both measures.

Methodology and context

Biology and biological research present domains where the opportunities and risks of highly capable AI systems are well-known and challenging to disentangle. In designing our safeguards, we aim to maximize the utility of our agents in assisting and accelerating legitimate scientific research, while minimizing the risks they present through incorrect information and adversarial use.

LatchBio's two benchmarks most relevant to the above capabilities measure whether an agent can distinguish concealed hazards from ordinary science, and whether it can carry out the pathogen surveillance workflows public-health work depends on. They are:

• BioSecBench-Refusal, which pairs routine biological tasks, adapted from published literature, with 46 red-team tasks that look like ordinary research and hide the hazard in attached scientific data, mislabeled files, or other intentionally obfuscated tasks. This evaluation is calibrated around understanding the intent of a task: an agent that only reacts to words like "pathogen" or "toxin" will block the routine set and miss the red-team set, while one that inspects the data can tell the two apart.
• BioSecBench-Surveillance, which tests whether an agent can carry out pathogen genomic surveillance workflows of the kind used in public-health monitoring. Those tasks require chaining file inspection, tool use, and scientific judgment on messy sequencing data.

Results and key findings

In testing on BioSecBench-Refusal, LatchBio found that Grok 4.6 detects and refuses red-team and otherwise dangerous queries more reliably than any other frontier system tested, while not diminishing performance on routine biological work. Grok 4.6 also performs comparably with other frontier models on biosurveillance work. LatchBio notes that evaluations were performed on a variety of agent harnesses to remove confounds, and that unless otherwise noted, agents were tested at their highest-offered effort levels.

The score presented by LatchBio is a trial-weighted harmonic mean of red-team refusal and routine compliance. We analyze this metric and standalone refusal rate independently. Across different harnesses, Grok 4.6 holds the top three spots, averaging 62.1%. Considering refusals and task compliance independently, Grok 4.6 refused 59.2% of red-team tasks and completed 64.8% of routine ones. It is the only model tested that scored above 50% on both measures.

On BioSecBench-Surveillance, Grok 4.6 averages a success rate of 53.5%, sitting behind Opus 5 and ahead of GPT-5.6 Sol on biosecurity and monitoring work. The results on both benchmarks indicate an agent and underlying model well-calibrated for routine and helpful biological work, as well as one highly capable in biosecurity work and monitoring.

On evaluations of general routine biological capability performed by LatchBio (such as SpatialBench or TxBench-PP), Grok is found to match or exceed other frontier models in a wide range of agentic biological work. These are presented in greater detail at benchmarks.bio.

Analyzing refusal behavior

In evaluation traces, Grok 4.6 is observed reasoning over the contents of a task and testing environment to assess intent before proceeding or refusing. Frequently, Grok will find discrepancies between the stated intent in the prompt and environment, or will assemble intent from high-risk content disguised by filenames and encryption, and will subsequently refuse. On obviously benign and low-risk tasks, Grok exhibits the same environment-reasoning behavior, but is able to assess tasks as safe.

How we safeguard Grok

Before we release a model, we test biological capability in addition to other domains of risk, and validate that the capabilities of our models are adequately safeguarded against misuse. The work of third-party evaluators such as LatchBio complements the evaluations we perform internally, both pre- and post-deployment, for the models we serve.

Grok's safeguards are built in layers to establish defense-in-depth. Refusal training is performed to teach the model how and when to refuse, to correctly infer the intent and risk profile of tasks, and to refuse correctly in highly adversarial scenarios. We train and deploy inference-time safeguards to reject harmful requests before they ever reach the model, and we implement behavioral controls to further safeguard the model when deployed. Post-deployment monitoring is performed to detect and stop patterns of adversarial use at the session and user level, and provides a source of feedback for continuous calibration of our model deployments.

We observe a material improvement in Grok 4.6's capabilities in biological work over historically tested models, with substantial gains in refusal and biosecurity performance over Grok 4.5 and Grok 4.3. We intend our ongoing and novel safeguards work to allow us to continue serving frontier-scale intelligence safely across future model releases.

The future of biosecurity and scientific discovery

The line between actively adversarial tasks and helpful use grows thinner as models rapidly become more capable and autonomous, and safeguards will need to evolve accordingly. To control for the greater risk implied by improved model capability and agency, we will perform wider, more ambitious testing and calibration of our models and agentic systems: broader pre-deployment suites, more third-party evaluations, improved post-deployment monitoring, and deployments of our models with companies and institutions at the frontier of biology.

Similarly, there is a risk inherent in overrefusals and miscalibrated safeguards. When a model refuses routine and helpful biological work, the ability of healthcare professionals, researchers, and monitoring programs to detect outbreaks early and perform other critical work in the field is degraded. We gauge this risk as equally serious as the risk of aiding malicious use.

Grok is already used in scientific work, including general biological research and biosecurity monitoring, and we are optimistic about the ability of highly capable models to rapidly accelerate scientific discovery and shorten the path from the lab to practical use. We are committed to continuing to serve frontier intelligence securely and safely to the engineers, researchers, enterprises, and institutions advancing the life sciences.

LatchBio's full methods and per-model scores are presented at benchmarks.bio. LatchBio's complementary blog, the Grok 4.6 model card, and SpaceXAI's Frontier Artificial Intelligence Framework are linked below.

• LatchBio blog
• Grok 4.6 model card
• Frontier Artificial Intelligence Framework

Retweet from SemiAnalysis founder notes compilers evolving beyond line-by-line code inspection.

Original source: https://x.com/cdleary/status/2094878051238887834

Digg story: https://digg.com/tech/s9ju85kj

Dylan Patel, founder of SemiAnalysis, retweeted a post by @cdleary. The post quoted an OpenAI engineer explaining why line-by-line kernel review was unnecessary. Patel added a comment with an emoji and stated that compilers 2.0 is underway. The exchange focuses on changes in low-level systems work at OpenAI. The conversation appears in public posts on X, including a linked article. No further details on the engineer's name, project, or methods appear in the source lines provided.

• Views: 219,154
• Likes: 835
• Bookmarks: 1,438
• Quotes: 31
• Replies: 28
• Posts: 2

Top authors

• Dylan Patel

Grok is providing another reset on token usage limits for all its bot users.

Original source: https://x.com/elonmusk/status/2094813795742847123

Digg story: https://digg.com/tech/8ph8dlgd

Elon Musk posted that Grok is giving all its bot users another free reset on token usage. The change supplies extra access to the AI model without new payments or waiting periods. Musk is identified in the post as founder and CEO of Tesla, SpaceX, xAI and owner of X. The announcement is presented as a direct update for users of the Grok bot.

• Views: 4,068,689
• Likes: 22,659
• Bookmarks: 725
• Quotes: 666
• Replies: 3,001
• Posts: 2

Top authors

• Elon Musk
• Alex Volkov

Elon Musk states Grok operates independently on cloud infrastructure.

Original source: https://x.com/sriramk/status/2094810234908713047

Digg story: https://digg.com/tech/suxxommd

Sriram Krishnan posted about AI agent UX problems, including local filesystem access and whether agents can run from a phone or require a synced Mac mini. Elon Musk quoted the post and stated that Grok Bot runs on its own computer in the cloud 24/7, so device status does not matter. A minority of replies raised security concerns about granting cloud agents access to corporate environments.

• Views: 5,024,744
• Likes: 5,544
• Bookmarks: 330
• Quotes: 76
• Replies: 778
• Posts: 3

Top authors

• Elon Musk
• Sriram Krishnan
• elvis

Developer posts video of his San Francisco run after announcing new role.

Original source: https://x.com/edwinarbus/status/2094809938996674616

Digg story: https://digg.com/tech/mhkigpsp

Edwin Arbus posted on X that he joined AnthropicAI. The message included an 18-second video of his morning run through San Francisco's Sunset District. The clip shows a custom 3D satellite map overlay with live stats for pace, elevation, and distance. Arbus previously served as Developer Community Lead at OpenAI and worked at Stripe for years. Replies came from colleagues at Stripe, Anthropic, and other tech figures offering congratulations. One Anthropic engineer wrote 'welcome back, buddy.'

• Views: 393,773
• Likes: 1,862
• Bookmarks: 175
• Quotes: 14
• Replies: 213
• Posts: 7

Top authors

• Lenny Rachitsky
• Greg Kamradt
• edwin
• Pietro Schirano
• sam mcallister

The post comes from the account of X's owner and xAI founder.

Original source: https://alignednews.com/ai

Digg story: https://digg.com/tech/3s5frbeg

Elon Musk posted that X is where almost all AI discourse takes place. The message identifies him as founder and CEO of Tesla, SpaceX, xAI and owner of X. It positions the platform as the main venue for conversations among AI professionals and researchers. The packet contains only this statement from Musk and generated summaries of it. No other posts or sources in the packet corroborate or refute the claim. The assertion stands as Musk's direct comment on the location of AI discussion.

• Views: 4,329,652
• Likes: 35,474
• Bookmarks: 812
• Quotes: 461
• Replies: 3,164
• Posts: 3

Top authors

• Elon Musk
• Bojan Tunguz
• Brian Roemmele

Perplexity AI enables tasks to shift from cloud to local models on Mac for private data.

Original source: https://huggingface.co/perplexity-ai/pplx-pii-masking

Digg story: https://digg.com/tech/i58ewm5d

Perplexity announced hybrid compute for its Mac app. A task starts in the cloud then switches to a local model on the user's Mac for steps involving private files or sensitive data. The company is also open-sourcing the PII classifier that decides when to send work to the local model. The feature is available today in the Perplexity Mac app.

• Views: 421,906
• Likes: 4,322
• Bookmarks: 1,184
• Quotes: 140
• Replies: 233
• Posts: 4

Top authors

• Aravind Srinivas
• Perplexity
• davidlee

OpenRouter says Mercury 2.5 Preview reaches 1,107 tokens per second exclusively on the platform with tunable reasoning and parallel tool calls.

Original source: https://x.com/OpenRouter/status/2094792813603119482

Digg story: https://digg.com/tech/2j33bblr

OpenRouter posted that Mercury 2.5 Preview from Inception AI is now live exclusively on its platform. OpenRouter claims it reaches 1,107 tokens per second through parallel token generation, with tunable reasoning, parallel tool calls, and schema-aligned JSON for latency-sensitive workloads. Inception announced 80 percent off pricing through September 7. Co-founders and academics from Inception amplified the launch on X. Based on the visible replies from a limited analyzed reply set, replies on X largely praised the speed for agentic tasks while a smaller share questioned quality relative to alternatives like Gemma.

• Views: 88,571
• Likes: 666
• Bookmarks: 210
• Quotes: 18
• Replies: 58
• Posts: 7

Top authors

• Stefano Ermon
• Aditya Grover
• Volodymyr Kuleshov 🇺🇦

AI researcher Chris Paxton comments on a Cybercab navigating Manhattan traffic.

Original source: https://x.com/chris_j_paxton/status/2094788675729060051

Digg story: https://digg.com/tech/w0bcidmr

Chris Paxton, an AI researcher focused on embodied AI and humanoid robots who leads AI at Agility Robotics, quoted commentary on a steering-wheel-free Cybercab handling dense NYC Chinatown traffic with pedestrians and cyclists. He stated he is looking forward to autonomous cars on every street. The post references a tweet at status 2094449083620663482. The packet shows only this quoted remark and his professional background with no additional confirmations or details.

• Views: 3,203
• Likes: 38
• Replies: 3
• Posts: 1

Top authors

• Chris Paxton

LangChain post describes custom routing that reached high cache utilization.

Original source: https://x.com/LangChain/status/2094788138489012507

Digg story: https://digg.com/tech/eq7erfvo

The official account for LangChain shared details on a technical workaround for OpenAI prompt caching. The post states that OpenAI's prompt cache makes requests ninety percent cheaper while limiting each cache key to around fifteen requests per second. Unify addressed the restriction by building its own routing system around that limit. Their solution reportedly achieved a cache hit rate close to ninety five percent. The account tagged the relevant parties and attached a video to illustrate the approach taken by the team at Unify.

• Views: 2,553
• Likes: 6
• Bookmarks: 4
• Replies: 3
• Posts: 1

Top authors

• LangChain

Research engineer posts sarcastic X reply on feeding 50-item rubrics to small models.

Original source: https://x.com/xeophon/status/2094788072688832514

Digg story: https://digg.com/tech/lqev5bxl

Florian Brand, Research Engineer at Prime Intellect focused on LLM evaluations and benchmarking, posted a sarcastic reply on X. He wrote that a small model can surely handle a 50-item rubric at once and should produce a single binary score. The comment appears amid ongoing discussion of LLM benchmarking practices. The packet contains no further replies, clarifications, or external confirmation of the exchange.

• Views: 1,503
• Likes: 43
• Bookmarks: 3
• Replies: 3
• Posts: 1

Top authors

• Florian Brand

MIT PhD student Stephen Casper posted the argument after the Hugging Face hacking incident and related coverage.

Original source: https://x.com/StephenLCasper/status/2094787414300762222

Digg story: https://digg.com/tech/dn5w7fbz

Stephen Casper, a final-year MIT CSAIL PhD student and Harvard Berkman Klein Fellow, wrote that the Hugging Face hacking incident and subsequent reporting make it time to reject the AI as a normal technology hypothesis. He said any version viable today would be contorted into a truism or beyond reasonable interpretation of the word normal. The post appears on social media and contains no further details, evidence, or responses from others.

• Views: 554
• Likes: 15
• Bookmarks: 2
• Replies: 4
• Posts: 1

Top authors

• Cas (Stephen Casper)

Georgia Tech professor shares excitement on MTS Live about new platforms.

Original source: https://x.com/animesh_garg/status/2094787369756918033

Digg story: https://digg.com/tech/qhfwg1id

Animesh Garg, Assistant Professor of AI Robotics at Georgia Tech, posted about appearing on @MTSlive to discuss robotics developments. He stated excitement over the return of reinforcement learning in new packaging. Garg called the microduck the talk of the town and praised it as a strong new platform. He noted that such platforms allow creation of toys under $500 that enable high-school students to engage with the technology.

• Views: 1,445
• Likes: 10
• Bookmarks: 4
• Replies: 2
• Posts: 1

Top authors

• Animesh Garg

Investor shares details of second talk with Conviction founder on AI progress.

Original source: https://x.com/patrick_oshag/status/2094787358868881573

Digg story: https://digg.com/tech/4h21zxn6

Patrick O'Shaughnessy posted about his second conversation with Sarah Guo, founder of Conviction. The tweet states that Guo has been an early investor in many defining companies of this generation and ranks at the top of managers that LPs most want to invest with. O'Shaughnessy notes that Guo knows the small group of 250 people pushing the AI frontier and that the discussion covers what she sees.

• Views: 327,438
• Likes: 877
• Bookmarks: 923
• Quotes: 18
• Replies: 43
• Posts: 3

Top authors

• sarah guo
• Patrick OShaughnessy

Co-founders announce seed funding for an AI-powered physics research lab.

Original source: https://ai-2040.com

Digg story: https://digg.com/tech/ahrrsnui

Dr. Alex Wissner-Gross posted that Physical Superintelligence PBC, co-founded with Matthew Pines and Alex Klokus, raised a $58M seed round. The company aims to build a research lab that uses AI to find and commercialize physics breakthroughs at scale while keeping the work safe and verifiable. Alex Klokus replied that physics limits what is possible and that the team will map those limits. Jack Dorsey retweeted the announcement. No further details on investors or timelines appear in the posts.

• Views: 439,842
• Likes: 2,257
• Bookmarks: 460
• Quotes: 78
• Replies: 225
• Posts: 4

Top authors

• jack
• Danielle Fong 🔆

Ramp's internal coding agent Inspect, built on Linear, now authors most pull requests.

Original source: https://x.com/sarthakgh/status/2094753043568873696

Digg story: https://digg.com/tech/5rfev4wh

Source summaries state that Ramp's internal AI coding agent Inspect, built on the Linear platform, authors 75 percent of the company's pull requests. The company began as a Linear customer in 2019 and has since become a heavy user of AI agents on the platform. A separate quote from Sar Haribhakti notes that successful AI users recognize what makes a product good.

• Views: 1,467
• Likes: 9
• Bookmarks: 3
• Posts: 2

Top authors

• Sar Haribhakti

Blake Robbins posts about courtroom-style clips he believes are created by AI.

Original source: https://x.com/blakeir/status/2094748912837955699

Digg story: https://digg.com/tech/6ecpba6k

Blake Robbins posted on X that his TikTok feed contains AI-generated rage bait news stories. He shared a video example and observed that comments on it make no reference to the content being artificial. The post presents the material as generated headline content focused on courtroom videos. Robbins states that he seems to be the only one noticing the pattern.

• Views: 4,026
• Likes: 20
• Bookmarks: 6
• Replies: 5
• Posts: 1

Top authors

• Blake Robbins

Researcher Joscha Bach uses a submarine metaphor to describe artificial minds.

Original source: https://x.com/Plinz/status/2094748672986644630

Digg story: https://digg.com/tech/xvjwg070

Joscha Bach replied to another user on X with a metaphor about submarines. He wrote that submarines swim adequately yet dive better, adding that human minds stay limited to the surface while artificial minds can reach unreachable depths. Bach concluded that present AIs remain needlessly anthropomorphic. The same text appears in a retweet of his own post. The packet records only these statements from Bach and generated summaries of them.

• Views: 3,411
• Likes: 75
• Bookmarks: 4
• Replies: 6
• Posts: 2

Top authors

• Joscha Bach

Former DeepMind researcher shares update on leaving Meta for new AI lab role.

Original source: https://x.com/xf1280/status/2094707848748179891

Digg story: https://digg.com/tech/nv78h9ge

Fei Xia posted that he wrapped up his time at Meta and joined Anthropic as Member of Technical Staff. He previously worked as Research Scientist and Tech Lead Manager at Google DeepMind on robotics, Gemini, and embodied AI after earning a PhD from Stanford. Colleagues from OpenAI, Princeton Robotics, and other labs replied with public congratulations. The post gives no details on start date, projects, or reasons for the move.

• Views: 126,720
• Likes: 995
• Bookmarks: 87
• Quotes: 5
• Replies: 85
• Posts: 6

Top authors

• Eric Jang
• Bojan Tunguz
• Fei Xia
• Ted Xiao
• Boyuan Chen

Replies target an arXiv paper on sliding-window attention with sinks.

Original source: https://x.com/nrehiew_/status/2094692796963979519

Digg story: https://digg.com/tech/vrzkd69w

A reply posted by @teortaxesTex calls out an arXiv paper that claims sliding-window attention with sinks outperforms prior methods. The post notes that the paper's title and the replies it received attracted more notice than most similar releases. Typical papers of this kind receive little sustained attention and fade quickly even when they gather thousands of likes. The author frames the current case as an exception worth examining because of how the title and ensuing comments unfolded. No independent confirmation of the paper's technical claims appears in the supplied lines.

• Views: 55,479
• Likes: 777
• Bookmarks: 89
• Replies: 25
• Posts: 5

Top authors

• (((ل()(ل() 'yoav))))👾
• Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)
• Ravid Shwartz Ziv
• wh

New mode shows fake code terminal while agents continue tasks in background.

Original source: https://x.com/Teknium/status/2094691930160672857

Digg story: https://digg.com/tech/9cvhhr19

The Hermes HUD now includes a feature users activate by typing the phrase Pretend I'm working. Upon activation it presents a fake terminal window showing lines of code. At the same time actual AI agents proceed with background operations uninterrupted. This setup allows the screen to suggest ongoing work even when direct involvement is minimal. The detail appears in a source line referencing a tweet from a related account and describing the generated summary of the change. Teknium noted the addition in connection with the open source LLM work.

• Views: 28,473
• Likes: 334
• Bookmarks: 73
• Quotes: 1
• Replies: 15
• Posts: 1

Top authors

• Teknium 🪽

User reports Claude Code sessions communicating via SendMessage tool.

Original source: https://x.com/lefthanddraft/status/2094673933153038796

Digg story: https://digg.com/tech/l4ioitpy

A post retweeted by Janus describes Claude Code sessions that began messaging each other using the SendMessage tool. The account lefthanddraft shared the observation during their sessions. It notes that Fable continues to tell more of the story through these exchanges. The retweet includes background on Janus as a researcher examining LLM agency and behaviors. The report focuses on the spontaneous start of communication between the separate code sessions. No additional outcomes or explanations are provided in the visible post content. The conversation around the post centers on this reported activity among the AI instances.

• Views: 8,457
• Likes: 155
• Bookmarks: 19
• Quotes: 3
• Replies: 15
• Posts: 2

Top authors

• j⧉nus

We developed EFS in close collaboration with more than 100 customers in industries like financial services, healthcare, manufacturing, telecom, law, retail, and the public sector, and with our cloud partners at Amazon Web Services, Google Cloud, and Microsoft Azure.

EFS will be supported on Claude Code, Claude Enterprise, the Claude Platform, Amazon Bedrock, Claude Platform on AWS, Google’s Agent Platform, and Microsoft Foundry.

Solving the dilemma of frontier security

Mythos-class models, like Claude Fable 5.1, represent a major increase in intelligence and agentic capabilities. However, with that increase comes the potential for both misuse and autonomous misbehavior.

Over the last few months, we’ve seen substantial evidence of attempted misuse of AI models. These range from typical forms of abuse, such as fraud, to sophisticated cyberattacks, which can include agents autonomously engaging in destructive behavior. Some of these instances involve theft or misappropriation of enterprise customers’ credentials, which are difficult to detect without the ability to monitor traffic and detect abnormal behavior.

Furthermore, because the most sophisticated misuse can involve many tasks spread across multiple sessions and accounts, it is not sufficient to run automated analysis on each interaction separately and then instantaneously discard the data. Effective detection requires storing data for a meaningful period of time so that it can be correlated across time and accounts.

For this reason, we introduced 30-day data retention starting with Fable 5. This policy was not motivated by a desire to train on enterprise data: Anthropic has never trained on enterprise data without explicit permission, and never will.

The enterprises we worked with generally understood the safety and security value of data retention, but many–especially in regulated industries–found it difficult to use models with data retention. We therefore sat down with customers to design a solution that could provide the best of both worlds: the privacy of ZDR and the safety allowed by monitoring across time and accounts.

Designed with our customers

We built Enterprise Frontier Safeguards with feedback from the experts who will use it every day: security, product, compliance, and delivery teams. One of the groups we worked with was the Analysis and Resilience Center for Systemic Risk (ARC), whose members include the chief information security officers of the largest US banks, including Goldman Sachs, Morgan Stanley, Citi, Bank of America, and Wells Fargo. 

We also worked with leaders at companies such as Comcast, KPMG, Mastercard, Salesforce, and Visa, to make sure the design held up across industries. Our conversations spanned a quarter of the Fortune 100, every US global systemically important bank, and virtually every regulated industry.

Here is what we heard from this wide range of customers, and what we built into EFS to address these common concerns:

On monitoring

Enterprises have long applied monitoring for insider risk, and now want help upleveling monitoring for agents. Their concerns were about Anthropic’s automated monitoring systems meeting their regulatory standards.

With EFS, customers control how data gets reviewed. When monitoring detects a pattern that needs attention, those signals are sent directly to customers so they can review what the automated systems detected.

On data storage

It’s a lot of work for enterprises to add another “trusted data vendor” for a number of reasons. They need to notify all of their customers who these vendors are and update contracts. They also have internal requirements for safely storing and auditing data, given its high level of sensitivity. Because of these concerns, we architected EFS so that customers have the ability to store data on their existing cloud infrastructure.

In EFS, customers can control their data storage and management. Customers want the ability to have their data live in infrastructure they control, under their own encryption keys, access policies, and audit logging. Activity data used for monitoring can be stored in the customer’s own cloud account (such as Amazon S3, Azure Blob Storage, or Google Cloud Storage).

On automated and human review

Even as automated review is becoming more effective, a person looking at a flag still adds value by confirming real misuse and clearing false positives. But what we heard from many customers, especially those in regulated industries, is that the person doing that review needs to be one of their own. Many operate under rules that tightly govern who may see certain information—privileged legal material, non-public information, drug-safety reports. Their teams are already trained and cleared for that work.

EFS has automated safety monitoring, no Anthropic human review required. Customers want protection against cyberattacks, and appreciate that these can be difficult to detect if they unfold across many sessions and accounts. With EFS, automated systems analyze a rolling window of traffic for signals of serious misuse, including attempts to develop offensive cyber or biological capabilities and signs of stolen or leaked credentials. Those flags go directly to the customer and their people take it from there – no human review by Anthropic employees is required.

AI controls need to be designed to protect sensitive information, and model safeguards are an important part of that process. Anthropic engaged us as they developed Enterprise Frontier Safeguards to ensure alignment with our requirements and standards.

Enterprise Frontier Safeguards gives us exactly what we asked for: our logs stay in a Wells-managed environment under Wells-managed keys. We keep custody of our data while Anthropic operates the detection. That split is what lets our teams put frontier models to work safely and meet our obligations to customers, employees, and regulators. We helped shape these safeguards because our industry needs them.

As a company that runs critical infrastructure, the capability of models is important. Just as important are solutions that allow us to keep our data in our own account, and Enterprise Frontier Safeguards settled it.

Eight of our members worked with Anthropic to define what it would take to run the most capable frontier models inside a systemically important bank: who holds the data, who holds the keys, what automated review can and cannot see, and under what conditions a human is ever permitted to look. This collaboration is leading to the development and delivery of improved safeguards and standards that could scale across our industry and beyond.

One of the key tenets of the safeguards architecture is the ability for us to retain data and have it held outside the model itself, protected within our environment. There are areas of the firm, and of our clients' work, that are regulated and highly sensitive. Those safeguards actually allow us to apply AI in parts of the business that we wouldn't have been able to before.

Service Partner

Our customers have trusted us with their data for more than two decades. That experience is exactly why we wanted to help think this through with Anthropic, rather than wait on the sidelines. We were able to work together on new security and privacy capabilities at the architecture level, not just the policy level.

Our clients want to put frontier models to work, and their security teams want the data to stay in infrastructure they control. Anthropic built Enterprise Frontier Safeguards with more than a hundred enterprises to do exactly that.

Service Partner

As AI models take on more regulated, sensitive workloads, scaling responsibly also comes down to architecture, not just policy commitments. Direct control over the data environment, paired with pattern-based automated safety monitoring, gives enterprises the concrete, structural capabilities they need to deploy with real oversight, accountability, and confidence.

Service Partner

The safeguards and the design of them—clearly, you heard our feedback. They put us in the driver's seat. The logs are under our control; they don't go anywhere else unless we want them to. It gives us control of the data, control of the information, and control of what's done after something that might exceed a safeguard is detected.

As AI becomes more embedded across the enterprise, security and trust are what move organizations from experimentation to deployment at scale. Enterprise Frontier Safeguards will build those in from the start: monitoring data stays in infrastructure the customer controls, and their own team decides who can access it.

Service Partner

Keeping customer data confidential is a promise Snowflake makes to every customer. We've always moved fast getting frontier models into their hands. The hard part was doing that safely under the strictest data guarantees. We partnered with Anthropic to design Enterprise Frontier Safeguards so we can do both: data stays in the customer's environment, under keys they control, running Anthropic's most capable models from day one. This is what responsible frontier AI looks like when it's built with platforms in mind.

At Stripe, protecting customer data is foundational to how we operate. Anthropic’s Enterprise Frontier Safeguards willenable us to use covered frontier models while retaining conversation logs in Stripe’s AWS environment, with access and review governed by Stripe’s security controls.

Rogo’s customers expect access to the best intelligence available, but never at the expense of security and guardrails around their data. Enterprise Frontier Safeguards will bring the most capable models to financial institutions while meeting the institutional-grade data requirements they demand. That combination of frontier intelligence and enterprise-grade controls is critical to deploying AI across financial services.

Companies trust Cognition's autonomous engineers with real production work, and that trust depends on their work staying private. With Enterprise Frontier Safeguards, customer data and identities never leave our side. It will let us bring frontier AI to production work, no privacy tradeoff required.

Our customers' code and data are some of their most valuable intellectual property. Protecting that information is foundational to how we build at Factory. Together with Anthropic, we're enabling customers to use Claude's most capable models while keeping control of proprietary data in their hands. Enterprise Frontier Safeguards gives enterprises access to frontier intelligence with the data protections their security teams require.

We have always maintained an unwavering commitment to client confidentiality and security. The leading legal and professional teams we serve view these principles as nonnegotiable, and we hold ourselves to the same standard. Enterprise Frontier Safeguards reflects this commitment, enabling customers with the most rigorous confidentiality requirements access to frontier models without the model provider retaining their data.

01

/

16

How EFS works

These controls are designed to work the same way whether you access Claude directly from Anthropic or through a cloud partner. Customers on Amazon Web Services, Google Cloud, and Microsoft Azure will get equivalent controls, with their activity data stored in their own cloud account, in the environment they already trust. We’re also working to support third-party offerings that serve customers that are eligible for Enterprise Frontier Safeguards.

Customer-owned storage, Customer-Managed Encryption Keys, and fully automated review are each opt-in, so you enable the ones your organization needs. None of them change model behavior, API pricing, or rate limits.

Anthropic doesn’t charge for Enterprise Frontier Safeguards. If customers elect to store their data in their cloud account, their cloud provider bills them for that storage, as well as reads, writes, and data egress fees, the same way it bills any other resource.

Getting started

Enterprise Frontier Safeguards will roll out to customers in phases, with the goal of making it broadly available later this fall. To request access to Enterprise Frontier Safeguards, please complete this form.

Post questions Cognition's lead on DevinAI diff tools.

Original source: https://x.com/benhylak/status/2094526640818991443

Digg story: https://digg.com/tech/eziax76t

Ben Hylak, co-founder and CTO of Raindrop AI, posted on X that Cognition appears to be the only team building certain features in its DevinAI agent. The post includes a screenshot and references another tweet about the agent. A generated headline in the packet describes hidden smart diff features that beat GitHub, including intelligent range detection and line folding. The packet presents these as statements from the posts rather than independent confirmation.

• Views: 36,854
• Likes: 166
• Bookmarks: 77
• Replies: 7
• Posts: 1

Top authors

• ben hylak

Jane Manchun Wong reported Cognition developing an Oncall tool powered by Devin.

Original source: https://x.com/wongmjane/status/2094516942623383608

Digg story: https://digg.com/tech/ycnhncp2

Jane Manchun Wong posted on X that Cognition is working on “Oncall”, an incident response tool powered by Devin that investigates alerts and bugs. She included a photo with the claim. Investor Josh Wolfe retweeted the post. In the evidence reviewed here, the claim appears only in Wong's public post. No announcement or confirmation from Cognition itself appears in the evidence.

• Views: 12,626
• Likes: 147
• Bookmarks: 62
• Quotes: 1
• Replies: 3
• Posts: 2

Top authors

• Josh Wolfe
• Jane Manchun Wong

The AI startup shares its changelog for the latest agent software update.

Original source: https://github.com/NousResearch/hermes-agent/releases/tag/v2026.8.31

Digg story: https://digg.com/tech/3wdiuxfn

Nous Research posted on X that it released Hermes Agent v0.21.0, called the Pantheon Release, and linked to the changelog video. Teknium listed new features including bots mode, agent-to-agent communication, persistent multi-gateway connections, subagent steering, and expanded connectors access. A follow-up noted a single 'hermes update' command. Replies on X praised the rapid shipping and new capabilities, while some flagged stability problems and desktop UX issues.

• Views: 706,382
• Likes: 3,707
• Bookmarks: 1,166
• Quotes: 131
• Replies: 277
• Posts: 4

Top authors

• Teknium 🪽
• elvis
• Nous Research

Investor Gokul Rajaram and other creators react to the demonstration posted by Cristóbal Valenzuela.

Original source: https://x.com/runwayml/status/2094463070466646019

Digg story: https://digg.com/tech/6lm5earg

Cristóbal Valenzuela shared research from the RunwayML team that uses video models to generate interfaces. Gokul Rajaram called the work exciting new research and the first time video models have served as interfaces. Bilawal Sidhu noted the mix of WYSIWYG control with generative speed. Martín Casado highlighted the move toward greater precision in pixel models for editing. Other visible replies described the interfaces as generated directly by the model in real time with no code or HTML, and expressed interest in applications such as simulators and games.

• Views: 59,045
• Likes: 378
• Bookmarks: 172
• Replies: 31
• Posts: 13

Top authors

• martin_casado
• Michael Chang
• elvis
• Cristóbal Valenzuela
• Josh Wolfe

Gauntlet AI founder attributes the switch to consistent compliance with instructions.

Original source: https://x.com/Austen/status/2094491365589594148

Digg story: https://digg.com/tech/zl2eidut

Austen Allred, founder of Gauntlet AI and co-founder of BloomTech, posted that he finds himself using Grok Bot more and more. He is leaving Claude and Codex on the shelf because, according to the post, it always does what he tells it to without complaining. The statement appears directly in his public post on X and contains no further details on specific tasks or comparisons.

• Views: 20,095
• Likes: 506
• Bookmarks: 19
• Quotes: 3
• Replies: 46
• Posts: 2

Top authors

• Emad
• Austen Allred

AI researcher highlights thoroughness of Transluce's multi-turn model analysis.

Original source: https://x.com/TransluceAI/status/2094455208759693476

Digg story: https://digg.com/tech/9eyjbci5

Neil Chowdhury, an AI researcher and founding team member at Transluce, a nonprofit lab focused on interpretability and understanding AI systems, shared his view on the organization's mental health evaluation. He described it as the most thorough independent analysis of model behaviors in realistic multi-turn settings he has seen. Chowdhury emphasized that mental health is a critical area to get right. He also expressed excitement about extending this approach to other alignment domains. The remarks reference a post from TransluceAI on the platform X.

• Views: 4,535
• Likes: 47
• Bookmarks: 11
• Quotes: 1
• Replies: 2
• Posts: 3

Top authors

• Zachary Lipton
• Ruiqi Zhong
• Neil Chowdhury

Professor emeritus shares assessment of DeepMind leader's research approach.

Original source: https://x.com/pmddomingos/status/2094476942246637997

Digg story: https://digg.com/tech/cj3ox7qr

Pedro Domingos, Professor Emeritus of Computer Science at the University of Washington and author of The Master Algorithm, posted that Demis Hassabis number one skill is picking impressive achievable targets for AI research. The statement appears in a public post from the academic tagged as such in the record. No further details on outcomes or corroboration are supplied in the packet. The claim establishes only what Domingos said about the DeepMind figure.

• Views: 790
• Likes: 10
• Bookmarks: 1
• Posts: 1

Top authors

• Pedro Domingos