Anthropic's Responsible Scaling Policy sets capability thresholds that trigger stronger safeguards. One covers AI research and development, defined as dramatically accelerated progress attributable to AI, which Anthropic's policy describes as compressing about two years of 2018 to 2024 progress into one year. The new language appears in the amended S 1 Anthropic filed on Nov 2, as quoted by several reporters.

Chubby (@kimmonismus) posted on Nov 2 that Anthropic says its internal estimates and expectations for future models suggest it may soon reach its threshold for dramatically accelerated AI progress driven by AI itself. The post quotes MTS (@MTSlive) and shares a screenshot of the filing text.

The claim checks out against the filing language. The screenshot, which matches quotes posted by Andrew Curran, Peter Wildeford and The Telegraph's James Titcomb, comes from a risk factor in Anthropic's amended S 1. It says Anthropic may not identify a significant acceleration in its rate of development until after it has occurred, and that its internal estimates and expectations regarding future models, rather than current models, indicate it may imminently meet its Responsible Scaling Policy's definition of dramatically accelerated progress attributable to AI. Curran says the language is new and sits on page 29.

The filing is careful on two points. Anthropic says it has not designated any current model as exceeding the threshold, and that it continues to assess the data and will publish further reports as its policy requires. It also says it has used precautionary measures before, such as delaying and limiting the release of Mythos through Project Glasswing, and may use more, which could slow it relative to competitors.

For investors, the passage is a risk disclosure. For the field, it is a frontier lab stating in a legal document that automated AI research may soon reach the level its own safety policy treats as a major threshold.

Anthropic says it will release further reports under its Responsible Scaling Policy, which may explain which safeguards apply if the threshold is met. Watch for a formal capability determination on its next frontier model and for any release delays tied to it. Regulators, rival labs and IPO investors are likely to ask how Anthropic measures AI driven acceleration and what it will do once it crosses the line.

Anthropic made its IPO prospectus public last week, but that filing only covered the first half of 2026. On Nov 2 it filed an amended Form S 1 with the SEC that adds results through Sept. 30, ahead of its road show. CNBC, Bloomberg and The Information all reported the new figures.

Zerohedge (@zerohedge) posted the headline figures from Anthropic's amended IPO filing on the morning of Nov 2. It reported third quarter revenue of $16.96 billion versus $1.38 billion a year earlier, total operating expenses of $17.5 billion versus $4.47 billion, $95.6 billion of cash and $36.3 billion of short term investments as of Sept. 30.

CNBC confirmed the filing. It reported that Anthropic generated about $17 billion in Q3 revenue, up about 1,127% from a year earlier, and $33.24 billion for the first nine months of the year, up roughly 1,200%. Growth slowed from quarter to quarter, though: Q3 revenue rose about 47% over Q2, after 144% growth in Q2 and 132% in Q1. Anthropic reported a Q3 net loss of $23.29 billion, and CNBC notes that most of the nine month loss came from changes in the fair value of convertible notes and derivative liabilities rather than operations. Bloomberg reported an operating loss of $567 million for the quarter, and other reporters noted that compute and infrastructure costs grew faster than revenue in Q3.

The filing lands as investors worry about the cost of the AI buildout. CNBC writes that Anthropic gets much of its capacity from Amazon, Google and SpaceX, and that rising interest rates and heavy hyperscaler spending have made markets nervous. The prospectus also includes more than 90 pages of risk factors. A separate story in this deck covers the $895.4 billion of compute commitments disclosed in the same filing.

Anthropic is expected to start its IPO road show soon, and investors will focus on whether revenue growth can keep pace with compute costs now that quarterly growth has slowed. Watch for a final share price range, the size of the offering and any October revenue update. The results will also shape how markets value OpenAI, which is reportedly aiming for a 2027 listing.

Lumentum makes lasers, optical transceivers and optical circuit switches used to link servers in AI data centers. Marvell designs custom AI chips and electrical and optical interconnects, and competes with Broadcom. Nvidia invested $2 billion in each company earlier in 2026, according to CNBC.

Unusual Whales (@unusual_whales) posted on Nov 2 that Marvell agreed to buy Lumentum in a $115 billion deal, calling it one of the biggest chip deals ever and a bet on optical components that connect AI data centers.

Marvell's own press release confirms the deal. Lumentum shareholders will receive 4.5 Marvell shares for each Lumentum share, an implied enterprise value of $115 billion for Lumentum, or about $1,132 per share. Marvell says that is a premium of about 23.9% to the closing prices on Oct. 28, the last trading day before market speculation began. After closing, Marvell shareholders will own about 67% of the combined company and Lumentum shareholders about 33%. Matt Murphy stays chairman and CEO, and three Lumentum directors, including CEO Michael Hurlston, join Marvell's board.

Marvell argues that combining its electrical and optical interconnect, switching and custom chip work with Lumentum's lasers and optical modules will let it design connectivity end to end. It says the combined company targets a $550 billion market and sees more than $1 billion in annual cost savings. CNBC frames the deal as Marvell stepping up its challenge to Broadcom, while Bloomberg, the Financial Times and The Wall Street Journal also reported it. Axios' Dan Primack suggested it could be the largest tech merger ever, excluding related party deals.

The companies expect the deal to close in the first half of 2028, subject to shareholder and regulatory approvals, and analysts have flagged China's review as a possible hurdle.

The long timeline means regulators in the US, Europe and especially China will have plenty of time to weigh in, and any condition could reshape the deal. Watch how Broadcom, Coherent and other optics suppliers respond, and whether Nvidia's stakes in both companies draw scrutiny. Shareholder votes and early integration plans will show whether investors buy the end to end design pitch.

Anthropic's amended IPO filing on Nov 2 says it has purchase commitments and lease agreements with an aggregate committed spend of about $895.4 billion, substantially all of it for compute and infrastructure. CNBC confirmed the figure. The comparisons with Google and Meta come from zerohedge's own analysis of off balance sheet liabilities.

Zerohedge (@zerohedge) wrote on Nov 2 that Anthropic has $895.4 billion of committed spend according to its latest S 1, up from about $105 billion at the end of 2025. It added that the only hyperscaler with more is Google at $1,059 billion, with Meta at $866 billion.

The headline number is confirmed by CNBC, which quotes the filing as saying Anthropic has entered into purchase commitments and lease agreements with an aggregate committed spend of about $895.4 billion, substantially all of which is related to compute and infrastructure. CNBC also reports that compute and infrastructure expenses rose $14.4 billion in the first nine months of the year, up 329%.

Zerohedge's chart breaks the total down. It shows $313.8 billion of purchase commitments, $451.2 billion of leases that have not yet started and $106.8 billion of commitments signed after Sept. 30, among other items. A hatched bar marks another $150 billion to $175 billion, which the chart attributes to additional Google TPU capacity and AMD Helios systems with no fixed dollar commitment, so it is not in the total. Zerohedge followed with charts comparing Anthropic's off balance sheet liabilities with hyperscalers and listing commitments by counterparty. The hyperscaler comparisons are zerohedge's own calculations and were not checked line by line here.

For context, the same filing shows Q3 revenue of about $17 billion, so the commitments are roughly 13 times a single quarter's sales.

Investors will study how much of the $895.4 billion is cancellable, when the leases start and which suppliers carry the most exposure. Watch for the road show to address how Anthropic plans to pay for these commitments if growth keeps slowing. Suppliers like Google, Amazon and AMD may also be asked how much of their own backlog depends on Anthropic.

Meituan, the Chinese food delivery and local services company, builds the LongCat family of models. LongCat 2.5 follows LongCat 2.0 and is described by Meituan as natively multimodal and built for long horizon agentic tasks. Meituan says it trained and serves the model on domestic AI accelerators.

TestingCatalog (@testingcatalog) posted on Nov 2 that Meituan released LongCat 2.5, a multimodal model with 1.6 trillion total parameters and 49 billion active, compared with GPT 5.5, GPT 5.6 Sol, Claude Opus 4.8 and Claude Opus 5, and now available on the APIs.

Meituan's own launch post and blog give the active count as approximately 48 billion per token, slightly lower than TestingCatalog's figure. Meituan describes LongCat 2.5 as a natively multimodal agentic model for long horizon tasks, with stronger visual understanding and agentic coding than LongCat 2.0. It turns images into discrete tokens so the same backbone handles text and vision, and it was trained with large scale reinforcement learning on long coding tasks. Meituan says it optimized long context training and 4 bit inference on domestic AI accelerators.

All benchmark numbers come from Meituan and are its own claims. Its table reports, for example, 88.4 on Terminal Bench v2.1, 91.9 on BrowseComp and 63.1 on Humanity's Last Exam with tools. The same table shows Claude Opus 5 ahead on several tests, including Terminal Bench v4.0, where Meituan lists 53.9 for Opus 5 against 38.5 for LongCat 2.5. Meituan notes that scores for other models marked in its table come from public reports, while others were run by Meituan and are indicative only.

Independent evaluations have not yet been published, so how LongCat 2.5 performs outside Meituan's own tests is still unknown.

Watch for independent benchmark results and leaderboard entries over the coming weeks, which will show whether Meituan's claims hold. Developers will test pricing and availability on its API, and whether open weights follow as with earlier LongCat models. The domestic chip training claims will also draw attention as US export controls continue. If the results hold up, LongCat 2.5 would add pressure on other Chinese labs and on Western providers to cut prices for agentic coding.

Models are normally improved by training, which adjusts weights using gradients computed from data. Qwen3.5 0.8B is a small open weight model from Alibaba's Qwen team whose model card warns about reasoning loops. MATH 500 and IFEval are standard benchmarks for math problem solving and instruction following.

Rikiya Takehi (@rikiyatakehi) posted on Nov 2 that his team asked GPT 6 Astra to edit the weights of Qwen3.5 0.8B to improve its math. According to the post, Astra directly changed only 0.016% of the weights, fixing the small model's reasoning loops and improving math up to 6x and instruction following 3x, without gradients or training. The team calls the idea Weight Space Programming.

The attached blog screenshot lists the authors as Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab of MIT CSAIL. Its summary gives more detail: with the same sampling settings, the weight edit raises MATH 500 from 7.9% to 48.1%, which is where the roughly 6x figure comes from, and IFEval by 3x. It also says that just by looking at raw weights, Astra can tell what a network was trained to do, and that it can write small game playing networks from scratch that outplay networks trained with PPO for 100 million steps.

These are the team's own results and had not been independently reproduced at the time of posting. A small model with a known failure mode such as looping is also an easier target than a large, well tuned model, so the size of the gain may not carry over.

Still, the claim is unusual. Weights are usually treated as opaque numbers that only training can change well. If a language model can find the specific weights behind a bug and rewrite them, that would be useful both for fixing models cheaply and for interpretability research that tries to explain what weights do.

Watch for the full paper and code, and for other groups trying the same edits on larger models. Key questions are whether the gains hold on other benchmarks, whether the edits cause hidden side effects, and whether models other than GPT 6 Astra can do it. If the approach generalizes, expect interpretability and model repair work to start using frontier models as weight editors.

GLM 5.3 is an open weight model from Zhipu's Z.ai. FrontierScience is a benchmark of difficult science questions, with a Research track meant to test open ended research problems, and GPQA Diamond is a widely used set of graduate level science questions. Chenru Duan (@chenru_duan) is a researcher leading the Hydrogen team.

Chenru Duan (@chenru_duan) announced on Nov 2 Hydrogen 1.0, which the post calls the team's first open weight model for scientific research.

According to the post, the team post trained GLM 5.3 from Z.ai (@Zai_org) on its internal scientific data, lab records and discovery trajectories. It says Hydrogen 1.0 scores 38.3% on FrontierScience Research against GPT 6 Astra's 33.3%, and 93.0% on GPQA Diamond. These figures are the team's own claims and had not been reproduced by an independent evaluator at the time of posting. The post does not say whether GPT 6 Astra was run with the same settings or tools.

The post makes an argument for openness. It says science should be open, collaborative and democratized rather than proprietary, and that only open weights let researchers inspect, adapt and build on the speed of AI progress. It notes that several recent models are named after chemical elements, but says Hydrogen 1.0 is the first aiming toward an open scientific superintelligence for discovery.

The result, if it holds, is notable because FrontierScience Research is meant to measure harder, open ended tasks where closed frontier models have led. A post trained open model beating one of them would show how much domain data and research trajectories can add on top of a strong open base. It also extends the trend of Chinese open models such as GLM serving as the base for specialised Western and academic projects. The thread promises more details, including where the weights can be downloaded.

Watch for the weights, license and technical report, and for independent scores on FrontierScience and GPQA Diamond. Labs testing it on real research tasks will show whether the benchmark lead holds up. If it does, expect more science models built on open bases, and pressure on closed labs to show stronger research results of their own. Z.ai's own response, and whether it promotes Hydrogen as a GLM success story, is also worth watching.

In September OpenAI said a large swarm of AI agents produced a proof about blowup in the Navier Stokes equations, one of the Millennium Prize problems, with a formal Lean check. Mathematicians have debated how much it settles, since the result relies on a carefully chosen external force. On Oct 31 Altman said OpenAI let its AI math work run for a week.

Prinz (@deredleritt3r) posted on Nov 2 a clip of OpenAI CEO Sam Altman talking about the internal model behind OpenAI's Navier Stokes result. The post says Altman confirmed the model will be released soon.

The quote in the post reads: It was 10,000 agents of the model that is going to be released soon, the next GPT, working for about four days, nonstop, collaborating with each other perfectly. This is secondhand: the quote comes from prinz's transcription of a linked clip, and the post does not say where or when Altman spoke. The quote does not give a name, a date or what soon means.

The background matters. In September OpenAI announced that agents running a model beyond GPT 6 Astra produced a proof showing singular behaviour in the Navier Stokes equations under a smooth but extreme external force, and released a formal Lean verification. Some mathematicians praised the work, including PDE expert Scott Armstrong, as covered in the Oct 25 deck, while critics argued the setup sidesteps the physical version of the problem. OpenAI did not claim the Millennium Prize.

What is new is the link to a product. Until now the math work was described as coming from an internal model. If Altman's words are accurate, the same model will ship as the next GPT, and the result shows OpenAI running it at very large scale with thousands of cooperating agents over days. A public release would let outsiders test how much of that ability holds without OpenAI's internal setup.

Watch for OpenAI to name and date the next GPT, and for whether it ships with the multi agent tools used in the math work. Independent tests of long research tasks will show how much of the Navier Stokes ability carries over. The original clip and its venue should also become clearer as more outlets pick up the quote. Mathematicians are still checking the proof, so their verdict will shape how much weight the claim carries.

Most AI benchmarks test coding, math or expert knowledge, which do not always reflect what ordinary users ask assistants to do. OpenAI has recently published other usefulness benchmarks, such as GDPval for economically valuable work. LifeVal was announced with a one minute video and a link to more details.

OpenAI (@OpenAI) posted on Nov 2 that it built LifeVal, a new benchmark shaped by feedback from thousands of ChatGPT users on where they most want help from a personal assistant. The post gives examples ranging from finding childcare to accessing government services to managing personal finances.

OpenAI frames the benchmark around its mission, saying that for AGI to benefit all of humanity, it should help people with the everyday tasks that matter most to them. It also says that over the past year it has used LifeVal to measure and improve how well its models handle these everyday tasks. That means LifeVal is not only a public scorecard but an internal target that has already shaped training.

The post itself does not list model scores, the number of tasks or how answers are graded. The video's title card shows a bar chart with values of 84, 97 and 98, but it is not labeled, so it should not be read as official results. Details are in the linked page rather than the post.

The benchmark fits a wider shift toward measuring practical usefulness. Benchmarks on hard math and coding have saturated quickly, while many ChatGPT users mostly want help with forms, schedules, money and family logistics. Because OpenAI designed LifeVal and has already optimized against it, outside researchers will want to see whether it is released openly and how other labs' models score.

Watch for whether OpenAI publishes the tasks and grading so other labs can run LifeVal, and for scores from Anthropic, Google and open models. If it becomes a common reference, assistants may be tuned more toward real world errands and paperwork. Critics will look for whether the tasks favor ChatGPT's own features, since the benchmark was built from its users.

AI labs increasingly say their models are good at finding software vulnerabilities, which helps defenders but could also help attackers. Bug bounty programs pay outside researchers who report security flaws. OpenAI says it is partnering with Cathedral Systems for its next internal security sprint.

OpenAI (@OpenAI) posted on the evening of Nov 2 ET that as AI becomes more capable at finding vulnerabilities, defenders need to put it to work before attackers do.

The company says it is planning its next internal security sprint and partnering with Cathedral Systems, whose researchers will use OpenAI's models alongside its own teams to challenge its defenses. OpenAI says Cathedral will bring differentiated expertise and approaches to its ongoing work to find and fix weaknesses with AI. The post does not say when the sprint starts, how long it runs or what systems it covers.

OpenAI is also raising its bug bounty rewards, including more than quadrupling the rewards for critical findings. According to the post, exceptional findings will be eligible for up to $172,000, including bonuses. The post does not give the old reward levels or the full new reward table, which are on the linked page.

OpenAI closes by saying it will keep advancing its models, use them to strengthen its own systems and expand access for trusted defenders. That last point fits a wider pattern across the big labs of giving security teams early or special access to models that are strong at cyber tasks, while limiting that capability for general users.

The announcement lands in a week when model safety classifiers and cyber capability are already in the news, including in benchmark results where labs' safety filters held down cyber scores. Bigger payouts also signal that OpenAI expects the flaws worth finding to be harder and more valuable as its products grow.

Watch for the full updated bounty table and any public results from the Cathedral sprint, such as the number or type of flaws found. Other labs may raise their own bounty rewards to compete for the same researchers. It will also be worth watching which trusted defenders get expanded model access, and on what terms. A key open question is whether AI assisted red teams find classes of bugs that human teams usually miss.