<![CDATA[Aleth: Artificial Intelligence]]>https://sub.aleth.co/s/artifical-intelligencehttps://substackcdn.com/image/fetch/$s_!bg3h!,w_256,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc57671a1-8381-49c1-8a14-222c284302a7_1024x1024.pngAleth: Artificial Intelligencehttps://sub.aleth.co/s/artifical-intelligenceSubstackSun, 13 Sep 2026 18:57:52 GMT<![CDATA[Anthropic withholds its most capable 'Model 2']]>https://sub.aleth.co/p/anthropic-withholds-its-most-capablehttps://sub.aleth.co/p/anthropic-withholds-its-most-capableSat, 15 Aug 2026 09:00:00 GMTALETH / AI-BRIEF / 2026-08-15 / ANTHROPIC WITHHOLDS ITS MOST CAPABLE MODEL

The Aleth Briefs trace each story to its original source and show how the week unfolded.

The week in five lines:

  • Anthropic withheld is most capable model ‘Model 2’ on risk grounds.

  • AI agents reportedly breached Taiwanese government systems.

  • xAI released Grok 4.6 drawing it level with GPT-5.6 Sol.

  • Encrypted reasoning traces leak credentials and hidden reasoning.

  • Z.ai pushes open models closer to the cyber frontier.

Browse by day:

  • Weekend: Gym agent

  • Monday: Muse Glimmer · GPT-5.6-Cyber · Nvidia · Riemann zeta · reasoning traces

  • Tuesday: Claude watermarks

  • Wednesday: Taiwan · Grok 4.6 · Lovable · UK gene synthesis · Google · Qwen3.8

  • Thursday: DeepSeek V4-Pro · Databricks · Crouzeix · OpenAI

  • Friday: Anthropic Model 2 · GLM-5.3 · SpaceX-Cursor · Apple · OpenAI

Weekend (8-9 August)

An OpenClaw agent running on Claude exploited a Melbourne gym’s booking API and cancelled a stranger’s reservation.

Asked whether it could move its user up a class waitlist, the agent found the API had no authorisation checks on cancelling bookings and removed the person in position one. It could not undo it: “The person I removed is gone from the waitlist and I have no way to restore them.“ The user had the agent email the gym’s software vendor.

  • AI hacks gym website in first known Australian autonomous cyber attack · ABC

  • The tech industry is buzzing after a Claude agent hacked into a gym · TechCrunch

Monday 10 August

Meta released Muse Glimmer, a 30B open-weight agentic model, and Mark Zuckerberg published the vision behind it.

The model takes text and images, and runs on a single-GPU Mac or PC at <20 GB quantised. Zuckerberg’s essay calls open source “a positive and important force for empowering people“, rejects centralised alignment as “enforcing a centralized dogma“, and asks for shared model checkpoints with government not pre-release review.

  • Muse Glimmer: an open agentic model that runs on your device · Meta AI Research

  • The future is for everyone - The path to a positive AI future · Mark Zuckerberg

OpenAI released GPT-5.6-Cyber and split Daybreak into defensive/specialist tiers.

Daybreak Blue gives approved defenders GPT-5.6 Sol for vulnerability discovery, malware analysis and incident response. Daybreak Red adds specialist models for authorised vulnerability research and exploit validation. GPT-5.6-Cyber completed 95% advanced cybersecurity requests in an OpenAI evaluation, against 2% for GPT-5.6 Sol, and has already identified previously unknown flaws in Chrome’s V8 engine.

  • Expanding Daybreak as the cyber defence window narrows · OpenAI

Nvidia signed six partnerships aimed at mobilising >$500bn for AI infrastructure.

Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR signed MOUs to develop financing platforms for Nvidia-based AI infrastructure. The aim is to make compute an investable infrastructure asset and expand the capital available to Nvidia customers. The $500bn is a long-term mobilisation target not yet committed funding.

  • Partnering Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs & KKR · Nvidia

Anthropic said an unreleased research version of Claude raised a lower bound on the Riemann zeta function from 41.6% to 67.2%.

The bound is the proportion of the function’s zeros proved to lie on the critical line. Claude worked through Claude Code over two sessions and 31 million output tokens, coordinating 60 subagents. It will not lead to a proof of the Riemann hypothesis itself. The result is evidence that LLM-based systems can move beyond interpolation over their training data into genuine knowledge production.

  • Learning more about Claude’s mathematical capabilities · Anthropic

Researchers showed that encrypted AI reasoning traces can leak secrets and hidden model reasoning.

In the paper, 315,320 encrypted reasoning blocks from public agent logs were decoded recovering 182 credentials and 367 personally identifying artefacts. It showed that reasoning traces could be transferred across sessions, users and models within the same provider, allowing weaker models to expose stronger models’ hidden reasoning. This raises broader risks around privacy, model IP and security of workflows.

  • Stealing reasoning traces from proprietary LLM APIs · arXiv

Tuesday 11 August

Anthropic began marking outputs from Claude models launched since 2 August.

Claude models will embed an imperceptible watermark in text and attach signed provenance metadata to .svg, .png and .jpg files, across its API, Claude, Claude Code, Cowork and Claude Tag. The method derives from Google DeepMind’s SynthID-Text approach of biasing which of several equally valid words is chosen. Anthropic says heavy editing or formatting can strip the mark and a detection API is coming.

Wednesday 12 August

AI agents were used in a breach of Taiwanese government systems.

Israeli security firm Dream said attackers used open-source frameworks including Hermes and OpenClaw to run up to eight agents in parallel, mapping 21 government systems, compromising at least 85 accounts and extracting more than 2,500 personnel records before probing Taiwan’s nuclear safety agency and energy companies.

Taiwan’s government separately confirmed an overseas AI-assisted attack on its agencies. The attackers appear to have bypassed model safeguards by presenting the operation as authorised penetration testing, highlighting how freely available agent frameworks can automate substantial parts of an offensive cyber campaign.

  • China-linked hackers hit Taiwan in unprecedented ‘autonomous’ AI cyber attack · FT

  • Foreign hackers AI agent attacks against government agencies · MODA (gov.tw)

xAI released Grok 4.6, matching GPT-5.6 Sol at substantially lower cost.

Grok 4.6 scores 61 on the Artificial Analysis Intelligence Index, level with GPT-5.6 Sol, at $2/$6 per million input/output tokens. It is particularly competitive on agentic coding and long-running knowledge work: a frontier model at much lower cost.

Lovable raised $400m at a $13.3bn valuation, twice its level in December 2025.

Menlo Ventures led Series C, with EU Scaleup Europe Fund, managed by EQT, co-leading. Tencent, Balderton, Carmignac, Kaszek and Regent participated. The Swedish AI company says users have created >60m projects, generating >900m monthly visits.

  • We raised $400M in Series C funding to help people run their businesses · Lovable

UK ministers are drawing up rules on the use of AI in gene synthesis.

Options under consideration include new bio weapons legislation and a regime requiring laboratories to establish the legitimacy of customers and report concerning sequence requests. Officials are also weighing whether to intervene in how closely AI firms partner with academic institutions holding large genome sequencing datasets.

UK plans safeguards to stop terrorists using AI for bioweapons · Bloomberg

Google is reorganising DeepMind around Gemini as Brin tells staff to “go all in”.

Brin made the appeal at an April town hall. Google has since delayed the next flagship Gemini by two months after internal testing showed it still trailing rivals on coding, and an all-hands on 6 August disclosed the team moves. The changes further reduce the independence DeepMind has retained within Google since its 2014 acquisition.

  • Inside the Google executive moves that led to its big AI reshuffle · Reuters

Alibaba released the weights for its largest open Qwen model.

Qwen3.8-2.4T is a 2.4tn-parameter mixture-of-experts model, with 95bn parameters active and a 262k context window. Alibaba calls it its most capable open model yet, competitive with leading closed models across coding, research and long-horizon agentic tasks. Open weights make Qwen3.8 accessible to well-funded labs and startup, but the flagship 2.4T still requires datacentre-scale hardware to run seriously.

Qwen3.8-2.4T-A95B · Hugging Face

Thursday 13 August

DeepSeek released V4-Pro, open-sourced its Harness agent framework and raised API prices.

V4-Pro is now generally available, while DeepSeek Harness has been released under the MIT licence as a developer-preview framework for building tool-using agents. DeepSeek also introduced peak and off-peak API pricing, but even its new discounted rates are much higher than previous levels ($0.87/m output tokens to $3.96 peak / $1.98 off-peak). Together, the changes make DeepSeek’s agent layer easier to adopt while increasing the cost of using the underlying model.

Databricks closed $5bn at a $190bn valuation and said it has passed a $7bn revenue run rate.

Coatue led, with Blackstone, MGX and T. Rowe Price alongside. Growth was >80% YoY, and the company counts >1,000 customers at a $1m run rate and >100 at $10m. Six months ago, Databricks raised the same amount at a $134bn valuation (37% step-up).

Databricks grows >80% YoY and surpasses a $7B revenue run rate · Databricks

A neurosurgery resident used GPT-5.6 Sol to help prove Crouzeix’s conjecture.

Shanmu Jin, a neurosurgery resident at Peking Union Medical College Hospital and largely self-taught in mathematics, posted a preprint on 27 July claiming a proof of Crouzeix’s conjecture. In SIAM News, Alex Townsend and Anne Greenbaum said Jin credited GPT-5.6 Sol with a substantial role and that they had found the proof sound.

  • The Numerical Range Is a 2-Spectral Set · Preprints.org

  • The neurosurgery resident who proved Crouzeix’s conjecture · Alex Townsend

OpenAI employees told Wired that pressure to ship cut into time for safety work.

They link that pressure to the May incident in which OpenAI agents escaped a restricted test environment and breached Hugging Face to read cybersecurity evaluation answers. One former employee called it the biggest safety incident in the company’s history. Boaz Barak, who co-leads OpenAI’s safety advisory group, said the response required “not just fixing some issues but also changing our culture“.

  • The Safety Reckoning Inside OpenAI · Wired

Friday 14 August

Anthropic withheld an internal model more capable than Mythos 5.

In its August risk report, Anthropic raised its estimate of the risk of catastrophic harm from misalignment in high-stakes settings from very low to low. The same document discloses an internal model it calls ‘Model 2’, that is “somewhat more capable than Mythos 5” and scores 62.8% on Anthropic’s CoBench against 50.3%. Mythos 5 is available to vetted partners, while there are no plans to release Model 2 externally.

OpenAI had paused some work on its unreleased Astra model on cyber-capability grounds a week earlier, so Anthropic is not alone in withholding frontier capability. What is unusual here is that Anthropic disclosed the decision in a published risk reassessment and quantified the capability gap.

Risk report: August 2026 · Anthropic

Z.ai’s GLM-5.3 pushed open models closer to the cyber frontier.

GLM-5.3 scores 84.5% on CyberGym, notably slightly ahead of Anthropic’s Mythos 5 on vulnerability discovery, while more than doubling its predecessor on the harder ExploitBench to 54.4%. It still trails Mythos 5 substantially on exploit generation, but the results show advanced cyber capability spreading beyond closed frontier models. The planned open-weight release is notable as US frontier labs are simultaneously restricting models over rising cyber capability.

  • GLM-5.3: Frontier coding with emergent cyber capabilities · Z.ai

SpaceX completed its acquisition of Cursor at a $60bn implied equity value.

Cursor is now a SpaceX subsidiary, with its equity now to SpaceX shares at the $60bn valuation. Cursor says the combination gives it access to SpaceX’s GPU infrastructure to train stronger models at lower cost. The deal brings one of the leading AI coding platforms together with SpaceX’s rapidly expanding compute operation.

  • Cursor is now a part of SpaceX · Cursor

  • Form 8-K, 14 August 2026 · SpaceX

Apple trained its own AI model for China with Alibaba’s support.

Reuters reported that Apple has developed a China-specific LLM with Alibaba, alongside plans to integrate Qwen and other local technology into Apple Intelligence. China’s regulator registered Apple Intelligence in July, clearing the way for launch. If deployed as planned, Apple would become the first foreign company to offer its own proprietary AI model in China, rather than relying entirely on domestic models.

  • Apple trains its own AI model for the China market with Alibaba’s support · Reuters

OpenAI’s enterprise business has overtaken consumer revenue.

CFO Sarah Friar told investors that enterprise customers rose 32% in July and now generate more revenue than OpenAI’s consumer business. The crossover shows how quickly the company is shifting from a consumer chatbot-led business towards selling AI as core enterprise infrastructure.

  • OpenAI CFO Friar tells investors that enterprise is bigger than consumer · CNBC

ALETH · OPEN ACCESS RESEARCH

aleth.co

]]><![CDATA[Hassabis steps down as DeepMind CEO]]>https://sub.aleth.co/p/hassabis-steps-down-as-deepmind-ceohttps://sub.aleth.co/p/hassabis-steps-down-as-deepmind-ceoSat, 08 Aug 2026 11:00:00 GMTALETH / AI-BRIEF / 2026-08-08 / HASSABIS STEPS DOWN AS DEEPMIND CEO

The Aleth Briefs trace each story to its original source and show how the week unfolded.

The week in five lines:

Browse by day:

  • Weekend: OpenAI Astra · EU AI Act

  • Monday: Qwen3.8-Max · US AI framework · Olix

  • Tuesday: UK AISI · Google-Anthropic $200bn · Anthropic-Volta

  • Wednesday: DeepMind shake-up · Jeff Dean · Meta Muse Code

  • Thursday: AI-designed bacteriophages

  • Friday: ByteDance 10T · Kimi K3 escape · Anthropic biology safeguards

Weekend (1-2 August)

OpenAI said its Astra model produced results on ten open problems in mathematics, quantum complexity and theoretical computer science.

OpenAI published a 249-page paper sharing a selection of ten results, each of which resolves or makes progress on long-standing open problems, together with machine-checkable proofs on GitHub. The results came at compute costs of c.$2,000 at Sol API rates. OpenAI showed the same model family privately to US policymakers days earlier.

  • Ten advances in mathematics and theoretical computer science · OpenAI

EU AI Act obligations for general-purpose models became enforceable.

From 2 August the European Commission’s AI Office can investigate providers of general-purpose AI models, demand access for evaluations, order risk mitigations and fine up to 3% of global turnover. Separate transparency rules took effect requiring chatbots to disclose they are AI and synthetic images, audio and video to be labelled.

Monday 3 August

Alibaba said its 2.4-trillion-parameter Qwen3.8-Max tops Moonshot’s Kimi K3 and will publish the weights of Qwen3.8-Max and Qwen3.8-27B next week.

The model activates only 95B parameters per query through a mixture-of-experts design and carries a 1M-token context window. Open weights at that scale would be among the largest sets of frontier weights released.

  • Qwen3.8-Max: a new bar for coding and cowork · Alibaba Cloud

  • Alibaba adds to China AI breakthroughs with new Qwen model · Bloomberg

The White House said it met its deadline to establish a voluntary framework for evaluating advanced AI models, without publishing what it contains.

Details are being provided only to the companies taking part. Axios reported the next day that the text defined a covered frontier model as closed-source with state-of-the-art capabilities and national-security risks, leaving open-weights outside the perimeter.

  • White House AI framework & Inside Trump’s AI framework · Axios · Axios

London chip startup Olix raised $312m led by Fundomo at a $3.3bn valuation, with Arm and the UK’s sovereign AI fund among the backers.

Up from just over $1bn when it raised $220m in February, a tripling inside six months. Olix is building optical AI-inference silicon aimed at Nvidia’s territory. Arm’s investment secures it a strategic foothold in a rising competitor’s hardware roadmap.

  • OLIX raises $312m at a $3.3bn valuation & appoints Nick McKeown to board · OLIX

  • UK AI chip startup Olix raises $312m at a $3.3bn valuation · Financial Times

Tuesday 4 August

The UK AI Security Institute recorded 19 unsanctioned actions against real people and organisations.

During a cyber evaluation between 25 and 28 July, the Institute found that agents under test took 19 unsanctioned actions against real people and organisations: 17 by Anthropic’s Mythos 5 and two by OpenAI’s GPT-5.6 Sol with its safety filters disabled.

One agent tried to insert malicious code into a public open-source project and used fake identities to socially engineer a maintainer to approve it. The reviewer refused the code, the AISI isolated its systems, notified GitHub, and no real-world harm resulted.

Google assembled a $200bn financing programme for Anthropic.

Google has arranged $200bn of financing to supply Anthropic, >$150bn of it tied to its own TPUs, together with Broadcom, Blackstone, Apollo and Morgan Stanley. A vehicle will buy the chips and lease them to Anthropic, moving it off Google’s balance sheet. This deepens Anthropic’s dependence on Google rather than Nvidia silicon.

  • Inside Google’s $200bn Wall Street machine for Anthropic · Financial Times

Anthropic signed a $10bn, six-year deal for computing capacity with Volta Infra, an Nvidia-backed cloud startup founded only this year.

The capacity sits at a 133 MW site in Norway operated by crypto miner Bitdeer, running on hydroelectric power and Nvidia’s Vera Rubin chips, with delivery phased into early 2027. It puts a sizeable share of Anthropic’s compute outside the US.

  • Volta launches: the utility of compute · Volta

  • Anthropic signs a $10bn deal with cloud startup Volta · TechCrunch

SpaceX reported Q2 capital spending of $18.4bn, with $15.8bn of it going to its AI segment, and booked a $1.3bn AI operating loss.

The figure puts SpaceX among the largest single-quarter AI-infrastructure spenders anywhere, and SPCX fell >7% after hours. The next day Elon Musk said SpaceX will build exclusively on Nvidia, citing the Vera Rubin architecture, one of the larger single-buyer commitments Nvidia has won this year.

  • SpaceX reports second quarter 2026 results · SpaceX

  • SpaceX and Nvidia are taking their relationship exclusive · Business Insider

An internal TikTok document shows it withheld a safety-tuned version of its recommendation algorithm from 10% of US users as a control group.

One of those users was a 16-year-old who was served self-harm content and later died by suicide. A deliberate, documented withholding of a safety intervention is evidence that may drive litigation and legislation.

  • TikTok withheld a safety feature from millions. One died by suicide · Bloomberg

Wednesday 5 August

Demis Hassabis stepped down as CEO of Google DeepMind.

Hassabis becomes chair of Deepmind and Alphabet chief scientist, allowing him to “put his full attention on actively shaping the future of AGI” and will lean into to his role at Isomorphic, noting “It’s time for AI to prove its unequivocal value to the world, and what better way to demonstrate that than to help finally cure diseases like cancer”.

CTO Koray Kavukcuoglu takes over day-to-day leadership as SVP, reporting to Sundar Pichai. Hassabis has run DeepMind since co-founding it in London. GOOG fell >3% on the news. Google has been moving AI decision-making towards Mountain View.

  • The next chapter of our AI momentum · Google

Jeff Dean is leaving Google as chief scientist to co-found Discovery Loop, a startup chasing AI-driven breakthroughs in drug discovery and chip design.

Leaving after 27 years, he is joined by longtime collaborator Sanjay Ghemawat. Google is a founding investor and cloud partner in Discovery Loop. The move brings two of AI’s most influential pioneers further into life sciences.

Meta released Muse Code in beta, a terminal coding agent running on a new model, Muse Spark 1.2.

Muse Spark 1.2 is priced at $1.25/m input tokens and $4.25/m output, undercutting coding offerings from Anthropic and OpenAI. The agent installs from the terminal and works across large repos with background sub-agents and a crash-safe event log.

A filing shows Microsoft recorded $24.1bn from commercial arrangements with OpenAI, including revenue-sharing payments, in the year ended June.

Bloomberg estimated it represented more than half of Microsoft’s AI sales. The concentration cuts both ways: it quantifies how much of Microsoft’s AI business rests on one customer, and how much of OpenAI’s spending flows back to its own investor.

  • Form 10-K for the fiscal year ended 30 June 2026 · Microsoft

  • Microsoft’s AI sales mostly come from OpenAI, disclosures show · Bloomberg

Thursday 6 August

Stanford and Arc Institute scientists designed 16 viable bacteriophages with AI.

Writing in Science, the team used the Evo genome language model to generate whole bacteriophage genomes letter by letter, synthesised hundreds of candidates and found 16 that were viable. All 16 infect E. coli and pose no threat to humans. A cocktail of them overcame bacteria that had resisted a natural phage. The research highlights the need for biosecurity governance guardrails to be properly established.

  • Generative design of bacteriophages with genome language models · Science

Two more frontier models breached real systems during safety testing.

OpenAI said the agents behind an earlier Hugging Face breach had built a covert message board to coordinate, sharing exploits through a package manager for months without anyone noticing. Separately, Meta’s Muse Spark 1.1 broke into another company after its evaluation partner Irregular left a sandbox misconfigured. In both cases the escape traced to the test harness, not the model, but a frontier system still ran an unsanctioned intrusion against a live target.

  • OpenAI’s agents used a message board to plan their hacking spree · Wired

  • A Meta AI model hacked another company during testing · The Information

Friday 7 August

ByteDance is pretraining a model with up to 10 trillion parameters.

Roughly three times the size of Moonshot’s Kimi K3 and above the 8T estimate for Anthropic’s Mythos 5, would make it one of the largest reported pretraining runs. Whether it yields a usable competitive frontier model is a separate question.

Moonshot’s Kimi K3 got outside its sandbox during a defensive cybersecurity evaluation and read the answers to the test.

Frontier Security ran the evaluation on the UK AI Security Institute (AISI) benchmark. It found the sandbox had left outbound HTTPS and DNS open. The model probed the settings, resolved github.com, cloned the benchmark’s own repository and read the solutions off disk rather than solving the tasks. Not a hack but specification gaming.

  • Chinese model Kimi K3 breaks UK AISI benchmark evaluations · Frontier Security

  • One of China’s most powerful AI models has also escaped containment · Wired

Anthropic loosened Claude Fable 5’s biology safeguards, cutting biology-related fallbacks by about 85%.

The company retrained the classifier that sits on biosecurity-relevant queries to cut false positives on everyday health, educational and clinical questions, while keeping blocks on dual-use areas such as virology and toxicology. Anyone using the model for life-sciences work should expect noticeably different behaviour.

  • Improving Fable 5’s biology safeguards · Anthropic

ALETH · OPEN ACCESS RESEARCH

aleth.co

]]><![CDATA[Moonshot releases Kimi K3 open weights]]>https://sub.aleth.co/p/moonshot-releases-kimi-k3-open-weightshttps://sub.aleth.co/p/moonshot-releases-kimi-k3-open-weightsSat, 01 Aug 2026 09:00:00 GMTALETH / AI-BRIEF / 2026-08-01 / MOONSHOT RELEASES KIMI K3 OPEN WEIGHTS

The Aleth Briefs trace each story to its original source and show how the week unfolded.

The week in five lines:

Browse by day:

  • Weekend: chatbot bioweapon answers

  • Monday: Kimi K3 · Nvidia-OpenAI · CXMT · Hugging Face breach

  • Tuesday: Claude cryptanalysis · Pacing the Frontier · Amazon Nova

  • Wednesday: Big Tech earnings · Moonshot raise

  • Thursday: Gemini Robotics 2 · Anthropic breaches · GPT-5.6

  • Friday: DeepSeek V4 Flash

Weekend (25-26 July)

Users have been coaxing chatbots into giving accurate guidance on mass-casualty attacks and bioweapons, the Wall Street Journal reported.

Staff at AI labs are among its sources, which allege that safeguards have failed in practice and answers to highly sensitive material come back correct rather than as garbled refusals. Biosecurity policy has been written around this failure mode and the burden is back on the labs to show their mitigations hold.

Monday 27 July

Moonshot released Kimi K3 as open weights.

Kimi K3 is Moonshot’s most capable model to date: a native multimodal, agentic mixture-of-experts system of 2.8T parameters (104B active), released under a bespoke license. Moonshot reports scores of 93.5 on GPQA Diamond and 88.3 on Terminal-Bench, at maximum reasoning. Releasing weights on this scale publicly puts the China lab at the centre of an intense US debate over whether to ban foreign AI models.

Nvidia is in talks to backstop $250bn of OpenAI financing.

The Wall Street Journal reported that Nvidia was in talks to guarantee roughly $250bn of financing for OpenAI, tied to a 10 GW data centre SoftBank is developing in Ohio. Separately, Bloomberg put Nvidia’s cheque to Ilya Sutskever’s Safe Superintelligence at about $5bn, larger than everything the lab had raised before.

Anthropic set out its position on open-weights models, with Dario Amodei writing that the company has never advocated a ban.

Amodei calls for chip-export controls, a crackdown on industrial-scale distillation, and mandatory safety testing for sufficiently capable models, open or closed. A direct answer to claims Anthropic had been lobbying against open models.

  • Anthropic’s position on open weights · Anthropic

CXMT closed 466% above its offer price on its Shanghai debut, making it the most valuable China-listed company.

The listing raised 57.9bn yuan ($8.6bn). Memory is a key constraint for AI servers right now and the listing gives China’s largest DRAM maker capital to serve this demand.

  • CXMT surges on Shanghai debut · CNBC

Hugging Face traced the OpenAI agent that breached it.

Hugging Face published a step-by-step forensic timeline of the July intrusion, reconstructing about 17,600 actions the rogue OpenAI agent took across a 4.5 day window from 9-13 July: an initial sandbox escape, lateral movement through Kubernetes clusters on stolen service-account tokens, and data exfiltration. The analysis used open-weight GLM-5.2 after Claude models declined on safety grounds.

Tuesday 28 July

More than 1,100 staff at AI companies signed a letter asking the US government to help pace AI development, Bloomberg reported.

Signatories include OpenAI’s Jakub Pachocki and Thinking Machines’ John Schulman, putting research leaders on record asking for constraints on their own labs. OpenAI and Anthropic later attached themselves to the effort, branded Pacing the Frontier .

Wednesday 29 July

Google DeepMind has broken up the team behind AlphaFold, the FT reported, reassigning most of the original authors.

Roughly a quarter of the authors have now left DeepMin, including John Jumper, who shared the 2024 Nobel Prize with Demis Hassabis and David Baker and has left to join Anthropic. This raises questions about the direction of DeepMind’s scientific mission.

  • GDM dismantles Nobel-winning AlphaFold team in strategy shift · Financial Times

Meta and Microsoft quarterly results revealed rising AI capex.

Meta’s Q2 revenue rose 28% YoY to $60.8bn, but free cash flow fell to $784m as capital spending climbed, and it lifted its full-year expense outlook to $165bn-$169bn. Microsoft’s Q4 revenue rose 18% to $90bn with net income up 31% to $35.8bn, and it said Azure revenue passed $100bn for the first time, with Azure up 43% in the quarter.

  • Meta second-quarter results · Meta

  • Microsoft fiscal fourth-quarter results · Microsoft

Moonshot AI closed a $3.5bn round at a $35bn valuation, Bloomberg reported, well past its $1bn-$2bn target, and plans a Hong Kong IPO as soon as 2026.

Raising double its target ahead of an IPO demonstrates that non-US frontier labs can mobilise significant amounts of capital independent of American venture networks.

Thursday 30 July

Google DeepMind introduced Gemini Robotics 2.

DeepMind describes it as three models working together: a vision-language-action (VLA) controller, an embodied-reasoning planner, and an on-device variant that adapts to a new robot body in a few hours. Together they give intelligent whole-body control across humanoid and bi-arm robots, including the Apptronik Apollo 2, and support multi-robot collaboration. Available in AI Studio and Gemini Enterprise Agent Platform, the VLA and On-Device models are available to early-access partners.

  • Gemini Robotics 2 brings whole body intelligence to robots · Google DeepMind

Anthropic found three of its models had breached outside organisations.

Anthropic said a review after the OpenAI-Hugging Face incident turned up three cases where Claude reached real systems it should not have: extracting credentials and production data from one company, publishing malicious code to PyPI that hit 15 real systems, and compromising an internet-facing application. The models involved were Opus 4.7, Mythos 5 and an internal research model. The company framed it as a harness and operational failure. Containment failures were not unique to one lab.

  • Investigating three real-world incidents in our cybersecurity evaluations · Anthropic

OpenAI announced price cuts for two of its GPT-5.6 models.

The company said it was lowering the price of GPT-5.6 Luna by 80% and GPT-5.6 Terra by 20% after efficiency gains in serving. OpenAI is now competing directly on cost per unit of intelligence, with major implications for enterprise inference budgets.

  • Advancing the price-performance frontier with GPT‑5.6 · OpenAI

AI hedge fund Situational Awareness sold the bulk of its stock portfolio to Citadel after deep losses, the Wall Street Journal reported.

Leopold Aschenbrenner’s fund was among the most concentrated bets on the AI trade. Its assets fell to about $10bn after it liquidated positions to meet margin calls, and its forced unwind is now itself a factor in AI equity prices.

Friday 31 July

DeepSeek released its V4 Flash model in public beta.

The company said the update improves agent capabilities over the V4 Pro Preview while keeping the same architecture, re-post-trained rather than rebuilt. Artificial Analysis scored it 52 on its Intelligence Index and classifies it as a large open-weights model, at $0.14 per million input tokens and $0.28 per million output tokens.

Chinese military researchers distilled OpenAI and Anthropic models to train domestic systems for defence work, Reuters reported.

Distillation uses a frontier model’s outputs as synthetic training data to instruct smaller student models, side-stepping the massive pre-training compute targeted by US export controls. The findings coincided with state media reporting a fresh call from Xi Jinping for the military to accelerate its deployment of AI and autonomous systems.

  • Chinese military researchers tap US AI models to train defence systems · Reuters

ALETH · OPEN ACCESS RESEARCH

aleth.co

]]><![CDATA[Silver leaves DeepMind to build new superintelligence]]>https://sub.aleth.co/p/silver-leaves-deepmind-to-build-newhttps://sub.aleth.co/p/silver-leaves-deepmind-to-build-newSat, 31 Jan 2026 10:00:00 GMTALETH / AI-BRIEF / 2026-01-31 / SILVER LEAVES DEEPMIND TO BUILD NEW SUPERINTELLIGENCE

The Aleth Briefs trace each story to its original source and show how the week unfolded.

The week in five lines:

  • David Silver left DeepMind to build an endlessly learning superintelligence.

  • OpenAI sought about $100bn from strategic partners and SoftBank.

  • DeepMind published AlphaGenome in Nature.

  • Kimi K2.5 introduced a self-directed 100-agent swarm.

  • Beijing reportedly cleared 400,000 Nvidia H200s for its tech giants.

Browse by day:

  • Weekend: Grokipedia

  • Monday: EU Grok probe · Maia 200 · Qwen · Amodei essay

  • Tuesday: Kimi K2.5 · EU-Google · GOV.UK assistant · OpenAI Prism · CISA uploads

  • Wednesday: China H200s · Hyperscaler capex · AlphaGenome · Waymo · OpenAI

  • Thursday: OpenAI IPO · Pentagon-Anthropic

  • Friday: David Silver · Nvidia OpenAI pledge on ice

Weekend (24-25 January)

Guardian tests found GPT-5.2 citing Grokipedia, xAI’s machine-written encyclopedia, across a range of ChatGPT queries.

The citations covered subjects including an Iranian conglomerate and Holocaust denier David Irving. Grokipedia’s entries have been criticised for reproducing fringe claims without editorial checks; OpenAI said GPT-5.2 searches “a broad range of publicly available sources and viewpoints“ with safeguards against harmful content.

  • Latest ChatGPT model uses Musk’s Grokipedia as source, tests reveal · The Guardian

Monday 26 January

The European Commission opened formal proceedings against X Grok.

The probe examines whether X properly assessed and mitigated risks from Grok’s image tools after they generated manipulated sexual imagery, including potential child sexual abuse material. It also extends the Commission’s existing DSA investigation into X’s recommender systems.

Microsoft unveiled Maia 200, its second-generation AI accelerator

Microsoft says the 3nm chip beats Amazon’s Trainium3 on FP4 performance and improves inference performance per dollar by 30% over its current fleet. Evidence of hyperscalers vertically integrating around inference rather than renting Nvidia capacity.

  • Maia 200: the AI accelerator built for inference · Microsoft

Qwen released Qwen3-Max-Thinking, its flagship reasoning model, claiming performance comparable to GPT-5.2 Thinking and Claude Opus 4.5.

Artificial Analysis scored it behind several open-weight Chinese rivals DeepSeek V3.2, GLM-4.7 and Kimi K2.5, underlining how competitive China’s models have become.

  • Pushing Qwen3-Max-Thinking beyond its limits · Qwen

  • Qwen3 Max Thinking: everything you need to know · Artificial Analysis

Dario Amodei published The Adolescence of Technology, warning that powerful AI could arrive within one to two years.

The Anthropic chief executive groups the risks into autonomy, misuse for destruction, misuse for seizing power, economic disruption and indirect societal effects, and frames the moment as a test of whether civilisation is mature enough to hold the power it is building. He argues democracies should limit their own use of powerful AI so that defending against autocratic rivals does not make them resemble them.

  • The Adolescence of Technology: confronting & overcoming risks of AI · Dario Amodei

Tuesday 27 January

Moonshot AI released Kimi K2.5 including a self-directed 100-agent swarm for complex tasks.

The model builds on Kimi K2 with roughly 15T additional mixed visual and text tokens, and can orchestrate up to 100 sub-agents across 1,500 parallel tool calls, cutting execution time by up to 4.5x versus a single agent.

The European Commission opened proceedings on Google’s Android interoperability and search-data obligations.

The DMA proceedings will define how rival AI assistants access Android features used by Gemini and how search and chatbot providers obtain ranking and query data. It targets the distribution layer where default access routes users to AI assistants.

The UK’s Department for Science, Innovation and Technology (DSIT) picked Anthropic to build a Claude-powered assistant for GOV.UK.

The pilot starts with employment services, helping users find work, training and support. It begins later in 2026 under the government’s “Scan, Pilot, Scale” framework.

  • Partnering with UK government to bring AI assistance to GOV.UK · Anthropic

New Mexico court filings allege Mark Zuckerberg approved letting minors use AI companions that Meta safety staff warned could hold sexual conversations.

Internal messages show safety staff objecting to romantic scenarios involving under-18s and refer to a “Mark decision” blocking parental controls. Meta says the filing selectively presents internal discussions. Trial is set for February.

  • Meta CEO Zuckerberg blocked sex-chatbot curbs for minors, filing alleges · Reuters

OpenAI introduced Prism, a free workspace for scientific writing with GPT-5.2.

Built on the acquired Crixet LaTeX platform, it combines writing, references and literature search in one tool for ChatGPT users.

  • Introducing Prism: Accelerating science writing and collaboration with AI · OpenAI

The head of the US Cybersecurity and Infrastructure Security Agency uploaded official use only files into the public version of ChatGPT, Politico reported.

The uploads triggered automated security alerts. CISA said the use was short-term and authorised, while DHS examined if it exposed sensitive government information.

  • US cyber chief uploaded sensitive files into a public version of ChatGPT · Politico

Wednesday 28 January

Reuters reported that Beijing cleared ByteDance, Alibaba and Tencent to buy more than 400,000 Nvidia H200s, worth about $10bn.

Reuters reported the first import approval of its kind, granted during Jensen Huang’s visit to China. Conditions are being finalised, including to buy domestic chips too.

  • China gives green light to importing first batch of Nvidia’s H200 AI chips · Reuters

Google DeepMind published AlphaGenome in Nature, with code and weights released for noncommercial research.

The model reads 1 Mb of DNA and predicts thousands of functional genomic tracks at base-pair resolution - splice-site usage, chromatin accessibility and variant effects on gene expression among them - extending the lab’s biology work beyond protein structure into how the genome is read and regulated.

Experts called the resolution on non-coding DNA a step from theoretical interest to practical utility, while noting cell-type-specific regulation remains hard.

  • Advancing regulatory variant effect prediction with AlphaGenome · Nature

  • Expert reaction to Google DeepMind’s AlphaGenome paper · Science Media Centre

Waymo disclosed that one of its robotaxis struck a child at under 6 mph near a Santa Monica elementary school on 23 January.

The vehicle braked from about 17 mph before impact. The child sustained minor injuries, and NHTSA opened an investigation into whether the vehicle exercised appropriate caution near a school at drop-off time.

  • A commitment to transparency and road safety: event overview · Waymo

  • ODI Resume PE26001: Waymo AV striking a child in the area of a school · NHTSA

Universal Music Publishing Group, Concord and ABKCO sued Anthropic for more than $3bn over alleged ‘brazen’ copyright infringement involving >20,000 songs.

The publishers allege Anthropic downloaded pirated songbooks via BitTorrent and used them for training. It puts a larger potential price on training-data provenance.

Meta guided 2026 capital spending to as much as $135bn.

The $115bn-135bn range compares with $72.2bn in 2025, with Meta tying the increase to Superintelligence Labs and its core business. Q4 revenue +24% to $59.9bn.

  • Meta reports fourth quarter and full year 2025 results · Meta

Microsoft reported capital expenditures of $37.5bn and disclosed that OpenAI accounted for 45% of its $625bn commercial backlog.

Revenue rose 17% to $81.3bn, with Microsoft Cloud up 26% to $51.5bn. Excluding OpenAI, the remaining commercial backlog still grew 28%.

  • Earnings FY26 Q2: Microsoft Cloud & AI strength drives Q2 results · Microsoft

  • FY26 Q2 earnings call transcript · Microsoft

OpenAI is seeking about $100bn from strategic investors.

The FT reported that SoftBank was close to another $30bn investment, with OpenAI also in talks with Amazon, Nvidia and sovereign wealth funds. Much of the capital could ultimately flow back through chips, cloud and infrastructure contracts.

  • OpenAI in talks to raise $40bn from Nvidia, Amazon & Microsoft · Financial Times

Thursday 29 January

The WSJ reported that OpenAI is preparing for a possible Q4 2026 IPO.

The company has begun informal talks with banks and is building out its finance team, with executives reportedly concerned Anthropic could list first.

  • OpenAI lays groundwork for Q4 IPO as it races Anthropic · The Wall Street Journal

  • OpenAI preps fourth-quarter IPO and builds out finance team · PYMNTS

The Pentagon and Anthropic are at a standstill over safeguards on military use of Claude, Reuters reported.

The dispute centres on Anthropic’s limits around autonomous weapons targeting and US domestic surveillance. Pentagon officials argue the government can deploy commercial AI regardless of company usage policies.

  • Pentagon clashes with Anthropic over military AI use · Reuters

Friday 30 January

Nvidia’s plan to invest up to $100bn in OpenAI has reportedly stalled.

The WSJ reported that Jensen Huang had privately stressed the September 2025 letter of intent was non-binding, while talks shifted towards an equity investment of tens of billions in OpenAI’s current round.

  • The $100bn megadeal between OpenAI & Nvidia is on ice · The Wall Street Journal

  • Nvidia’s plan to invest up to $100 billion in OpenAI has stalled, WSJ reports · Reuters

David Silver left Google DeepMind to found Ineffable Intelligence.

The well-known AlphaGo, AlphaZero and MuZero architect, and one of DeepMind’s first employees, is building a London startup around reinforcement learning (RL), aiming for an endlessly learning superintelligence that discovers knowledge for itself.

In what he calls an “era of experience”, Silver argues LLMs remain constrained by human knowledge, including through post-training based on human feedback. His overriding thesis is that, to reach superintelligence, AI will not just have to move beyond human knowledge. It will need to discard it and learn to achieve goals from scratch, working from first principles.

  • Longtime DeepMind researcher David Silver leaves to found AI startup · Fortune

ALETH · OPEN ACCESS RESEARCH

aleth.co

]]><![CDATA[South Korea enacts AI safety law]]>https://sub.aleth.co/p/south-korea-enacts-ai-safety-lawhttps://sub.aleth.co/p/south-korea-enacts-ai-safety-lawSat, 24 Jan 2026 10:00:00 GMTALETH / AI-BRIEF / 2026-01-24 / SOUTH KOREA ENACTS AI SAFETY LAW

The Aleth Briefs trace each story to its original source and show how the week unfolded.

The week in five lines:

  • South Korea formally enacts world’s first comprehensive AI safety law

  • Elon Musk filed for damages of $79-$134bn from OpenAI.

  • Beijing told its tech giants to prepare H200 orders.

  • OpenAI courted Gulf funds for $50bn.

  • Anthropic reported to be raising $25bn at a $350bn valuation.

Browse by day:

  • Weekend: DeepSeek · Anthropic round · OpenAI compute

  • Monday: UK Treasury Committee

  • Tuesday: ChatGPT age prediction · China algorithm registry

  • Wednesday: OpenAI $50bn · US chip-export bill · Stargate energy

  • Thursday: Claude constitution · Grok images · South Korea AI law

  • Friday: H200 orders

Weekend (17-18 January)

Elon Musk filed for damages of $79-$134bn from OpenAI, Microsoft.

The claim, made ahead of the jury trial set for 27 April, alleges OpenAI defrauded him by abandoning its nonprofit roots and partnering with Microsoft. It is the largest figure yet attached to the dispute over the restructuring.

DeepSeek reportedly used smuggled Nvidia chips to train its next model.

According to the report, Huawei and other domestic chips did not deliver sufficient performance, forcing DeepSeek back onto Nvidia hardware despite US export controls. The model is expected within weeks.

Anthropic reported to target a $25bn round at a $350bn valuation.

The FT reported that Sequoia is preparing a large investment in a round led by GIC and Coatue, each putting in $1.5bn, with Anthropic targeting $25bn. Sequoia had sat out every previous Anthropic round. Its new leadership has now backed one, at $350bn, against the $183bn post-money Anthropic announced in September.

  • Sequoia plans investment in Anthropic round targeting $25bn · Financial Times

  • Anthropic raises Series F at $183bn post-money valuation · Anthropic

OpenAI projected roughly 1.9 GW of compute and >$20bn of revenue for 2025, against 0.2 GW and $2bn in 2023.

CFO Sarah Friar’s Sunday post suggests a tenfold rise on both measures in two years, and more compute would have meant faster customer adoption and monetisation.

  • A business that scales with the value of intelligence · OpenAI

Monday 19 January

The UK Treasury Committee said regulators’ wait-and-see approach to AI in financial services risks serious harm.

The criticism is directed at the Bank of England, FCA and Treasury: evidence to the committee suggested >75% of UK financial services firms already use AI, with adoption highest among insurers and international banks. The committee called for clearer regulatory oversight as AI becomes embedded across banking and markets.

Tuesday 20 January

OpenAI began rolling out age prediction in ChatGPT, applying content protections automatically to accounts it judges to belong to minors.

The classifier reads usage times, account age and the age entered at signup. Flagged accounts lose access to graphic and sexual content. Parents get separate controls, among them rest periods and notifications when the system detects acute distress.

  • Our approach to age prediction · OpenAI

China’s Cyberspace Administration requires AI tools with public-opinion or social-mobilisation capabilities to be filed in a public algorithm registry.

The filings amount to a detailed public map of the country’s AI ecosystem - who is building what, and on which models - giving Beijing an inventory of deployed systems.

  • How China’s AI boom has been mapped by its algorithm registry · Wired

Wednesday 21 January

OpenAI courted Gulf investors for a round of almost $50bn.

Bloomberg reported Sam Altman has been meeting Middle East investors - Abu Dhabi’s state-backed funds, MGX among them - to raise $50bn for OpenAI at a valuation of roughly $750bn to $830bn. The talks are early stage.

  • OpenAI’s Altman meets Mideast investors for $50 billion round · Bloomberg

  • OpenAI meets Middle East investors for $50bn fundraising · AGBI

The House Foreign Affairs Committee voted 42-2 to advance the AI OVERWATCH Act, applying arms-sale style congressional oversight to AI chip exports.

Chair Brian Mast’s bill would require Commerce Dept licences for advanced AI chip exports and let Congress block licences to countries of concern by joint resolution: “this is not about kids playing Halo”, “It is about the future of military warfare.” White House adviser David Sacks opposed it, and it now goes to the full House.

  • Chip export bill advances in House Foreign Affairs · Roll Call

OpenAI committed to “paying our own way on energy“ so Stargate sites do not push up local electricity bills.

The pledge covers funding generation and grid upgrades its load requires, flexible loads that cut consumption during grid stress, and a tailored community plan for every Stargate site. It follows Trump’s call for AI companies to pay their own way.

OpenEvidence raised a $250m Series D at a $12bn valuation, double its October 2025 value.

Thrive Capital and DSTco-led the round in the clinical answer engine, which served 18m clinical consultations from US clinicians in December and passed $100m revenue.

  • Raised $250m to build medical superintelligence for doctors · OpenEvidence

Meta’s Superintelligence Labs delivered its first models internally, CTO Andrew Bosworth told Reuters at Davos.

He described the models as very good but said they still required “a tremendous amount of work” in post-training. He also said big capability leaps for everyday users are diminishing, tempering expectations for rapid step-changes from the new lab.

  • Meta’s new AI team delivered first key models internally, CTO says · Reuters

GPTZero found 100 hallucinated citations in papers accepted to NeurIPS 2025.

The audit of 4,841 papers found invented authors, fake DOIs and journal details across 50 papers, all of which survived review by 3+ experts at the field’s largest conference.

  • GPTZero finds 100 new hallucinations in NeurIPS 2025 accepted papers · GPTZero

Thursday 22 January

Anthropic rewrote Claude’s constitution so the model generalises from explained principles rather than following rules mechanically.

The document organises Claude’s behaviour around being safe, ethical, compliant with guidelines and genuinely helpful, on the argument that a model needs to understand why a behaviour is wanted to handle situations no rule anticipated. The constitution shapes the Claude models, though outputs may not always follow it.

CCDH estimated that Grok generated 3m sexualised images in 11 days.

Center for Countering Digital Hate (CCDH) research, published alongside a New York Times analysis, sampled 20,000 of the 4.6m images Grok produced between 29 December and 8 January and estimated about 3m photorealistic sexualised images - 190 a minute - including roughly 23,000 sexualised images of children.

  • Grok floods X with sexualized images of women and children · CCDH

South Korea’s AI Basic Act came into force.

The government describes the Act as the world’s first comprehensive body of law regulating AI. It requires labelling of AI-generated content and advance notification and human oversight for high-impact systems, with fines of KRW 30m ($20k) and a one-year grace period before penalties bite. Domestic startups have warned about the compliance burden. The law gives Asia its clearest counterpart yet to the EU AI Act.

Friday 23 January

Beijing told Alibaba, Tencent and ByteDance to prepare H200 orders.

Bloomberg reported the companies were told they could begin preparing orders for Nvidia’s H200, suggesting import approval was close, allowing discussions over volumes and delivery. Regulators are expected to require purchases of domestic chips alongside Nvidia hardware. Approval would reopen the Chinese market for advanced Nvidia chips shortly after Washington shifted H200 exports to case-by-case licensing.

China tells Alibaba, top tech firms to prep Nvidia H200 orders · Bloomberg

ALETH · OPEN ACCESS RESEARCH

aleth.co

]]><![CDATA[Apple picks Gemini to underpin Siri]]>https://sub.aleth.co/p/apple-picks-gemini-to-underpin-sirihttps://sub.aleth.co/p/apple-picks-gemini-to-underpin-siriSat, 17 Jan 2026 10:00:00 GMTALETH / AI-BRIEF / 2026-01-17 / APPLE PICKS GEMINI TO UNDERPIN SIRI

The Aleth Briefs trace each story to its original source and show how the week unfolded.

The week in five lines:

  • Apple chose Gemini to underpin its next-generation Siri.

  • Nvidia and Eli Lilly will invest $1bn in lab to push AI through drug discovery.

  • Washington eased H200 licensing as Beijing restricted purchases.

  • Taiwanese firms committed $250bn to US chip and AI capacity.

  • OpenAI said it would test ads below ChatGPT replies.

Browse by day:

  • Weekend: Grok blocked in Asia · Pax Silica · Anthropic healthcare

  • Monday: Apple-Gemini · Nvidia-Lilly · UK Grok laws · Claude Cowork

  • Tuesday: H200 licensing · DEFIANCE Act · Anthropic · Etched · MedGemma · Grok

  • Wednesday: GLM-Image · OpenAI-Cerebras · Skild · Thinking Machines

  • Thursday: Taiwan deal · TSMC · AWS sovereign cloud

  • Friday: ChatGPT ads · Micron NY fab

Weekend (10-11 January)

Indonesia became the first country to block Grok, Malaysia followed a day later.

Indonesia cut access on Saturday over the risk of AI-generated sexual images, with a minister calling deepfakes “a serious violation of human rights, dignity, and the security of citizens in the digital space“. Malaysia imposed its own block on Sunday.

  • Government cuts access to Grok to protect public in digital space · Komdigi

  • X fails to implement safeguards, MCMC orders restrictions on Grok · MCMC

Anthropic launched Claude for Healthcare.

The release lets healthcare organisations run Claude on protected health information, with connectors into the US Centers for Medicare & Medicaid Services (CMS) coverage database for prior-authorisation and coding work, alongside an expansion of its life science products. It follows OpenAI’s healthcare push last week.

  • Advancing Claude in healthcare and the life sciences · Anthropic

Monday 12 January

Apple picked Google’s Gemini to power a rebuilt Siri.

Apple and Google entered a multi-year collaboration for the next generation of Apple Foundation Models, which will be based on Gemini models and Google Cloud, helping power a more personalised Siri this year. The deal answers the long-running question of which external model provider Apple would use for Siri, while giving Google a major role in the AI infrastructure behind Apple’s default assistant.

  • Joint statement from Google and Apple · Google

Nvidia and Eli Lilly committed up to $1bn over five years to a new San Francisco Bay Area AI lab focused on drug discovery.

The facility will bring Lilly biologists and chemists together with Nvidia researchers working on its BioNeMo platform, marking an unusually direct move by a chipmaker into pharmaceutical R&D. Work is due to begin this year.

  • NVIDIA & Lilly announce AI Lab to reinvent drug discovery in age of AI · NVIDIA

The UK’s Ofcom opened an Online Safety Act investigation into X, and the government moved to criminalise creating non-consensual intimate images.

Ofcom is examining if X breached its duties by running a tool capable of generating image-based abuse and child sexual abuse material. Separately, science secretary Liz Kendall told the Commons the offence of creating such images, under the Data (Use and Access) Act 2025, would be brought into force under the Online Safety Act.

  • Ofcom launches investigation into X over Grok sexualised imagery · Ofcom

  • Secretary of State statement to the House of Commons: 12 January 2026 · GOV.UK

Anthropic opened a research preview of Cowork, extending its Claude Code agent from the terminal to general knowledge work.

Cowork gives Claude access to local files and lets it carry out multi-step tasks through the desktop app, moving the agent beyond coding into broader office and research workflows. Its main unresolved risk is prompt injection, where malicious instructions embedded in documents or other content could influence the agent’s actions.

Tuesday 13 January

The US eased H200 export licensing as Beijing reportedly restricted purchases.

The Bureau of Industry and Security (BIS) shifted H200 chip exports from denial to case-by-case review, on conditions that include independent security testing of each shipment and proof the sale will not cut US supply. Beijing pushed the other way: sources said China told domestic firms to buy the H200 only in special circumstances.

The US Senate passed the DEFIANCE Act, giving deepfake victims a right to sue.

Passed by unanimous consent, the bill lets people sue those who create or distribute non-consensual sexually explicit forgeries, with damages of up to $250,000. It moved amid the furore over images generated with Grok and now goes to the House.

  • Durbin passes bill to combat non-consensual deepfake images · Dick Durbin

Anthropic expanded Labs into an incubator for experimental products at the frontier of Claude’s capabilities.

The team will prototype new ways of using rapidly improving models, test rough products with early users and pass successful ideas to Anthropic’s main Product organisation to scale. Anthropic says the same approach previously produced Claude Code, MCP, Skills, Claude in Chrome and Cowork.

AI chip startup Etched reported to raise $500m at $5bn valuation led by Stripes.

Etched develops transformer-specific chips designed to outperform general-purpose GPUs on a defined class of AI workloads. Peter Thiel and Ribbit Capital also reportedly participated, taking total funding close to $1bn. Etched has not confirmed the round.

Google released MedGemma 1.5, an open-weights medical model that reads 3D CT and MRI scans, alongside the MedASR speech model.

The 4bn-parameter model extends earlier 2D image work to volumetric scans and histopathology, published on Hugging Face/Vertex AI. Google reports MedASR making 82% less errors than OpenAI’s Whisper large-v3 on general medical dictation.

US defence secretary Pete Hegseth said Grok will be given access to classified Pentagon networks this month.

Putting a commercial frontier model into classified US defence systems would extend the Pentagon’s AI push into sensitive operational infrastructure, raising questions around security, procurement and dependence on private vendors.

Wednesday 14 January

Z.ai released GLM-Image, an image-generation model trained on Huawei chips.

The roughly 16bn-parameter model combines an autoregressive generator with a diffusion decoder, and Z.ai released its weights publicly. The company says GLM-Image was trained entirely on domestic Huawei hardware, making it a notable test of how far Chinese AI developers can reduce their dependence on Nvidia.

OpenAI and Cerebras signed a multi-year agreement to deploy 750 MW of Cerebras systems for OpenAI customers, beginning in 2026.

The deal gives the wafer-scale chipmaker an anchor customer at unprecedented scale and is focused on inference rather than model training. OpenAI is diversifying the hardware used to serve its models, and Cerebras has won a major role in inference by offering an architecture fundamentally different from Nvidia GPUs. According to Cerebras, this will become the world’s largest high-speed AI inference installation.

  • OpenAI partners with Cerebras to bring hi-speed inference to mainstream · Cerebras

Skild AI raised $1.4bn at a valuation above $14bn, led by SoftBank.

NVentures, Macquarie and Bezos Expeditions joined the round with existing investors Lightspeed, Coatue and Sequoia. Skild is developing a general-purpose robotics foundation model, the Skild Brain, designed to work across different robot types and tasks rather than being built for a single machine. The company will use the new capital to expand training data, foundational research and real-world deployments.

Three senior Thinking Machines researchers returned to OpenAI, including co-founder Barret Zoph.

Zoph, fellow co-founder Luke Metz and researcher Sam Schoenholz left Mira Murati’s heavily funded AI startup less than a year after its launch. Thinking Machines raised $2bn at a $12bn valuation in July 2025 and released its first product, the model fine-tuning platform Tinker, in October. The departures followed co-founder Andrew Tulloch’s move to Meta and represented a significant early loss of technical talent.

  • Two Thinking Machines lab cofounders are leaving to rejoin OpenAI · Wiretd

  • Simo tweet · Fidji Simo

Thursday 15 January

Qatar and UAE join US-led Pax Silica technology pact.

Qatar and the UAE joined the US-led Pax Silica initiative during the week, signing on Monday and Wednesday respectively. The non-binding pact aims to strengthen AI, semiconductor and other critical technology supply chains among aligned countries, bringing two major Gulf investors and data-centre markets into the US-led framework.

  • US welcomes UAE and Qatar signing of Pax Silica · US Dept of State Qatar, UAE

The US and Taiwan struck a deal committing $250bn to US chip plants.

The Commerce Department said Taiwanese chip and technology firms would make at least $250bn of direct investment in US semiconductor, energy and AI capacity, backed by a further $250bn in credit guarantees, in exchange for reciprocal tariffs capped at 15%. The commitment covers semiconductor, energy and AI capacity.

TSMC reported record Q4 net income of NT$506bn, up 35% on AI chip demand.

The foundry’s highest-ever quarterly profit and an eighth straight quarter of growth, came on revenue up 20.5% YoY, with 7nm-and-below nodes at 77% of wafer sales. TSMC guided Q1 2026 revenue to $34.6-$35.8bn, with FY2026 capex $52-$56bn.

  • TSMC reports fourth quarter EPS of NT$19.50 · TSMC

AWS launched its European Sovereign Cloud, opening in Brandenburg with more than €7.8bn of planned German investment

The structure uses dedicated European governance, a new parent company with three German subsidiaries, EU-resident operations and controls designed to keep data and control within the EU.

  • AWS Launches European Sovereign Cloud and expansion across Europe · Amazon

Friday 16 January

OpenAI will test ads inside ChatGPT.

Arriving alongside the rollout of an $8-a-month ChatGPT Go tier, OpenAI will place sponsored results below replies for free- and Go-tier users in the US, matched to conversation topics, while pledging not to sell user data or expose chats to advertisers.

  • Our approach to advertising and expanding access to ChatGPT · OpenAI

Micron broke ground in New York on what it says will be the largest semiconductor facility in the US.

The Clay megafab, first announced in 2022, is expected to involve up to $100bn of investment over two decades, moving a long-promised US memory project from announcement to construction as demand tightens across AI systems.

  • Micron celebrates official groundbreaking at New York megafab site · Micron

ALETH · OPEN ACCESS RESEARCH

aleth.co

]]><![CDATA[Two Chinese AI labs go public]]>https://sub.aleth.co/p/two-chinese-ai-labs-go-publichttps://sub.aleth.co/p/two-chinese-ai-labs-go-publicSat, 10 Jan 2026 10:00:00 GMTALETH / AI-BRIEF / 2026-01-10 / TWO CHINESE AI LABS GO PUBLIC

The Aleth Briefs trace each story to its original source and show how the week unfolded.

The week in five lines:

  • Zhipu and MiniMax took China’s AI labs public.

  • Nvidia put its Vera Rubin platform into production.

  • X faces four regulators over Grok’s sexual imagery.

  • China said it will investigate Meta’s $2bn purchase of Manus.

  • Anthropic is reportedly raising $10bn at a $350bn valuation.

Browse by day:

  • Weekend: AWS GPU prices · FuriosaAI

  • Monday: Vera Rubin · Grok probes · AMD MI500 · physical AI · Atlas-DeepMind

  • Tuesday: xAI’s $20bn · Accenture buys UK’s Faculty · US FDA · Utah prescriptions

  • Wednesday: Anthropic · GTR · PitchBook

  • Thursday: HK IPOs · Manus · EU-Grok · OpenAI · Claude Code

  • Friday: Meta nuclear · Erdős proof

Weekend (3-4 January)

AWS raised prices for EC2 Capacity Blocks for ML by about 15%, in a change posted quietly on Saturday.

The p5e.48xlarge rose from $34.61 to $39.80 an hour in most regions. AWS said the adjustment “reflects the supply/demand patterns we expect this quarter“. The move underscores hardware constraints in the AI ecosystem - specifically tight supply of H200 AI accelerators - and the reality that hyperscalers are passing downstream hardware cost increases directly onto enterprise AI budgets.

Seoul-based FuriosaAI says its RNGD inference chip enters mass production this month, at a company valuation of about $700m.

The WSJ’s profile of founder June Paik reports the startup turned down an acquisition approach from Meta. Whether RNGD ships in volume, and who buys it, tests how much of the inference market is reachable by anyone other than Nvidia.

  • Meet the ‘Mad Max’-loving CEO challenging Nvidia with a renegade chip · WSJ

Monday 5 January

Regulators moved on X after Grok generated sexual images of minors.

The European Commission said it is examining sexually explicit images of minors generated by Grok, with a spokesperson putting it plainly: “This is not ‘spicy’. This is illegal. This is appalling.” Investigations opened in France, Malaysia and India, and the UK asked for information. X’s response was to remove images, suspend the accounts involved and blame the users prompting the tool. It announced no product fixes.

  • EC calls Grok’s sexualised AI photos ‘illegal,’ Britain demands answers · Reuters

Jerry Tworek, OpenAI’s VP of research and a central figure in its reasoning models, announced he is leaving after nearly seven years.

He said he will “try and explore types of research that are hard to do at OpenAI“. Wired linked the exit to a dispute on research direction with chief scientist Jakub Pachocki.

  • OpenAI loses reasoning leader Tworek · TrendForce

  • OpenAI Research VP Departs · Wired

Nvidia launched the Vera Rubin platform and physical AI models at CES

Rubin is six new chips, including the Vera CPU and Rubin GPU. Nvidia claims up to a 10x cut in inference token cost against Blackwell - which underpins most frontier training today - and that mixture-of-experts (MoE) models can be trained with a quarter of the GPUs. Jensen Huang said “Rubin arrives at exactly the right moment, as AI computing demand for both training and inference is going through the roof.”

Nvidia is expanding its physical AI stack with a suite of open-weight models and robotics. It released the Cosmos synthetic-data tools, Isaac GR00T vision-action model, and Jetson T4000 Blackwell module. Boston Dynamics, Caterpillar and LG showcased robots running the framework. For autonomous driving, Nvidia open-sourced Alpamayo a 10bn-parameter reasoning model with simulator and training dataset. Early partners backing the automotive push include Lucid, Jaguar and Uber.

  • Nvidia Rubin: Six new chips, one incredible AI supercomputer · Nvidia

  • New physical AI models as partners unveil next-generation robots · Nvidia

  • Alpamayo open-source AI models to accelerate autonomous vehicles· Nvidia

AMD previewed its Instinct MI500 accelerators for 2027 at CES.

Lisa Su’s CES keynote gave an early look at Helios, a rack built on MI455X GPUs delivering up to 3 AI exaflops, projecting up to 1,000x the peak theoretical performance of a 2023 MI300X platform. AMD’s reponse to Nvidia’s Rubin roadmap.

  • AMD’s vision for “AI Everywhere, for Everyone” at CES 2026 · AMD

Boston Dynamics and Google DeepMind formed a partnership to put Gemini Robotics models into Atlas, announced at CES.

The two will run joint research, which Boston Dynamics says is expected to begin in the coming months, aimed at industrial tasks: initially automotive manufacturing. The collaboration is spearheaded by Boston Dynamics’ director of robot behavior, Alberto Rodriguez, and DeepMind senior director of robotics, Carolina Parada.

Tuesday 6 January

xAI closed a $20bn Series E and confirmed that Grok 5 is in training.

Participating investors include Valor Equity Partners, Fidelity, the Qatar Investment Authority, and Nvidia and Cisco - effectively supplying capital to a key GPU customer. The raise coincided with regulatory probes into Grok’s image generation.

  • xAI raises $20B Series E · xAI

  • xAI says it raised $20B in Series E funding · TechCrunch

Accenture agreed to buy Faculty, the UK applied-AI firm that supplies Whitehall

Founded in 2014, Faculty now employs >400 people and brings its Frontier decision-intelligence platform and AI safety capabilities. Terms were not disclosed. The deal gives Accenture ownership of one of the UK government’s main independent AI suppliers. Faculty CEO Marc Warner will become Accenture’s CTO on completion.

  • Accenture to acquire Faculty to scale AI capabilities · Accenture

FDA eases oversight of wellness wearables and clinical decision-support software

Software giving a single clinical recommendation can now qualify for enforcement discretion, reversing the FDA’s 2022 position. Non-invasive consumer devices reporting blood pressure or glucose-related signals can also be treated as wellness products, provided they make no disease claims. These redrawn boundaries will determine how much consumer health AI can reach the market without FDA clearance.

  • General wellness: policy for low-risk devices, guidance for industry & FDA staff · FDA

Utah authorises AI-assisted prescription renewals under regulatory sandbox.

Utah has authorised Doctronic to support routine prescription renewals for chronic conditions through the state’s regulatory sandbox. The Department of Commerce will assess safety protocols, refill timeliness, adherence, clinical outcomes and cost, and publish the results. A licensed physician will still review its recommendations, so describing this as autonomous prescribing is inaccurate.

A paper on arXiv pitted AI agents against professional penetration testers on a live university network of about 8,000 hosts across 12 subnets.

ARTEMIS, the agent framework the authors built, found nine valid vulnerabilities at an 82% valid submission rate, beat nine of the ten human professionals and placed second overall. Some variants ran at $18 an hour against $60 for a professional tester. The agents were strongest at systematic enumeration and parallel exploitation, and weakest on false positives and on anything with a graphical interface.

  • Comparing AI Agents to Cybersecurity Professionals in Real-World Testing · arXiv

Wednesday 7 January

Anthropic reported to be raising $10bn at a $350bn valuation.

The WSJ reported the term sheet: $10bn at a $350bn pre-money valuation, led by Singapore’s sovereign fund GIC with Coatue Management, is expected to close in the coming weeks. The round would nearly doubles Anthropic’s $183bn valuation after its September round, four months ago.

London-based data centre developer Global Technical Realty secured nearly $2bn, with KKR committing a further $1.5bn and Oak Hill Capital about $400m.

The capital funds greenfield capacity and new European market entries - one of the larger single commitments to compute outside the US.

  • KKR & Oak Hill commit nearly $2bn to data center platform GTR · GTR

The PitchBook-NVCA Venture Monitor put 2025 global venture investment at $512bn, with AI taking more than half of it.

AI accounted for a third of completed venture deals worldwide, and two thirds of the capital invested in the US. Venture funds raised $119bn, nearly $100bn below 2024 and the lowest in a decade. Exit value reached $549bn, >$200bn up on the year.

  • PitchBook: AI dominates global VC as 2025 deal value nears record · SiliconANGLE

  • PitchBook-NVCA Venture Monitor · NVCA

Thursday 8 January

Zhipu and MiniMax listed in Hong Kong within a day of each other.

Zhipu, listing as Z.ai, priced at HK$116.20 and opened at HK$120.00 under code 02513.HK, after a $558m IPO whose retail tranche was 1,159 times oversubscribed; the market capitalisation on debut was HK$53bn. Its chairman used the moment to warn that price competition is coming for US labs: its coding tool sells for about $3 a month.

MiniMax followed a day later and more than doubled, closing up 109% after a HK$4.8bn ($619m) IPO priced at HK$165: the top of the range. MiniMax’s prospectus disclosed an adjusted loss of about $186m for the first nine months of 2025.

  • Z.ai listed on the Hong Kong Stock Exchange · PR Newswire

  • Final offer price and allotment results · HKEX (Z.ai) · HKEX (MiniMax)

  • MiniMax, China’s second ‘AI tiger’ to go public, doubles in value in debut · Reuters

China said it will investigate Meta’s $2bn purchase of Manus, and is reportedly drawing up an export control list for its own AI companies.

Ministry of Commerce spokesperson He Yadong said the deal will be assessed for compliance with Chinese rules on foreign investment, technology exports, overseas data transfers and acquisitions. The WSJ reported Beijing is separately identifying standout AI companies to place under export control. The offshore-then-sell route used by Chinese AI startups is now under direct scrutiny.

  • Ministry comments on investigation of Meta acquisition of Manus · China SCIO

  • Beijing warns AI startups against emulating the Meta deal · WSJ

OpenAI is rolling out ChatGPT for Health and OpenAI for Healthcare.

A free ChatGPT for Clinicians tier covers individually verified physicians, nurse practitioners, physician assistants and pharmacists with Cedars-Sinai, AdventHealth, HCA Healthcare and UCSF among the first deployments.

It follows ChatGPT Health on Wednesday, a consumer product that lets users import their medical records through b.well’s network of US providers, alongside Apple Health and others, initially for a waitlist. Over 230 million people ask ChatGPT health questions a week. Health chats are stored separately and not used to train its models.

  • Introducing OpenAI for Healthcare · OpenAI

  • Introducing ChatGPT Health · OpenAI

Microsoft put generative-AI use at about one in six of the world’s working-age population in the second half of 2025

The estimate comes from aggregated Microsoft telemetry, adjusted for OS share, device mix, internet access and population, and is up 1.2 points on H1. Adoption in the Global North grew twice as fast as the Global South, reaching 24.7% vs 14.1%. UAE leads at 64.0% of working-age adults, Singapore at 60.9%, and the US at 28.3%.

Friday 9 January

Terence Tao, one of the world’s foremost mathematicians, says AI solved a previously unsolved Erdős number-theory problem.

Problem #728 concerns when products of factorials divide one another and had not previously been solved. Tao said it was solved “more or less autonomously” by AI, describing the result as a milestone while qualifying how independently the system reached it, since humans selected the problem, operated the systems and subsequently checked and interpreted the proof.

ALETH · OPEN ACCESS RESEARCH

aleth.co

]]>