DISPATCH

Everything that mattered in AI, one page a week.

Most AI news does not survive the week. This is the part that did — the releases, the research, and the shifts that actually change how we build. Designed & built to keep you up to date with things in AI without needing to be unemployed. Just refresh Saturday morning and review the last week's dispatch.

90 DISPATCHESWRITTEN EVERY FRIDAYNEXT UPDATE IN —

DISPATCH 06

WEEK OF FEB 1 – 7, 2025

Congress moved to ban a model it cannot un-publish

An MIT-licensed reasoning model spread across Hugging Face in hundreds of community distills — and the legislative response was to ban an app. Meanwhile GitHub turned Copilot into a coding agent.

Bipartisan legislation to purge DeepSeek from US federal devices was introduced, targeting an app built on MIT-licensed weights that anyone can self-host. Hugging Face's Open R1 project published its first verified reproduction of DeepSeek's reported MATH-500 scores.

Agent tooling moved fast: OpenAI shipped the deep research agent and GitHub put agent mode plus its autonomous SWE agent on the roadmap. Alphabet and Amazon guided to a combined $175B of 2025 capex, and the EU published its prohibited-AI guidelines the same day Google deleted its weapons and surveillance pledges.

FRI · Feb 7, 2025open weightsdeepseekregulationexport controlsreasoning

Washington moves to ban the open-weight reasoning model it can't un-publish

Bipartisan legislation to purge DeepSeek from US federal devices was introduced: H.R.1121, the No DeepSeek on Government Devices Act, filed and referred to the House Oversight Committee. It would direct OMB to write standards requiring removal of the DeepSeek app from federal agency IT, and extends to any successor product built by the hedge fund that funds DeepSeek.

It follows a Senate bill that reporting said would attach jail time to downloading DeepSeek. The irony is the target: DeepSeek-R1 is an MIT-licensed, openly downloadable reasoning model, shipped with six open dense distills. This week the community reproduction wave proved the point — Hugging Face's Open R1 project reproduced DeepSeek's reported MATH-500 scores for the entire R1-Distill family, and had to build a public eval leaderboard because the models' roughly 6,000-token average responses were too long to evaluate casually.

Washington moves to ban the open-weight reasoning model it can't un-publish — Hugging Face

WHY IT MATTERS

Open weights reached frontier-adjacent reasoning quality and the response was legislative, not technical — that is the trade a self-hosted stack now lives in. A frontier-adjacent reasoning model that anyone can host is simultaneously a national-security item in Congress. Also a reminder that the derivative ecosystem (third-party distills, quantizations, community evals) is where an open model's real value compounds.

Congress.gov — H.R.1121, No DeepSeek on Government Devices Act Hugging Face — DeepSeek-R1 model card Hugging Face — Open R1: Update #1 (reproduction of MATH-500 scores)

THU · Feb 6, 2025agentscoding agentsdeveloper toolsharness

GitHub turns Copilot into a coding agent and teases Project Padawan

GitHub shipped the agentic era of Copilot: agent mode in VS Code iterates on its own output, recognizes and fixes errors, suggests terminal commands and does runtime-error self-healing; Copilot Edits went generally available for multi-file natural-language edits; next edit suggestions, shareable prompt files and screenshot-to-UI-code arrived in preview.

Gemini 2.0 Flash and OpenAI's o3-mini entered Copilot Chat in public preview with org-wide model controls, and GitHub gave a first look at Project Padawan — an autonomous SWE agent that takes an assigned issue and returns work for review.

GitHub turns Copilot into a coding agent and teases Project Padawan — images.ctfassets.net

WHY IT MATTERS

The harness, not the model, is the product: self-healing loops, multi-file edits, prompt files and admin-level model selection are exactly the permission and configuration surfaces an agent builder has to secure. Prompt files and stored instruction sets also become supply-chain targets the moment they are shared — which is the TrapDoor attack class, two years early.

GitHub newsroom — agent mode, Copilot Edits GA, Project Padawan InfoWorld — GitHub Copilot previews agent mode

THU · Feb 6, 2025computecapexdata centershyperscalers

Amazon plans ~$100B of 2025 capex for AI, after Alphabet's ~$75B

On its Q4 2024 earnings call Amazon said it expects about $100 billion of capital expenditures in 2025, with Andy Jassy calling AI a once in a lifetime opportunity and the vast majority of the spend going to AWS AI capacity.

Two days earlier Alphabet guided to roughly $75 billion of 2025 capex, largely for AI infrastructure, alongside Q4 revenue of about $96.5 billion — a quarter that slightly missed expectations and dragged shares. Two hyperscalers, $175B of committed 2025 spend between them, in the same week the industry debated whether cheap Chinese reasoning models made that spend unnecessary.

WHY IT MATTERS

Capex guidance is the clearest leading indicator of inference cost and GPU availability: what gets built in 2025 sets the price floor agent workloads pay in 2026, and the DeepSeek cost-shock debate is exactly the argument this spending is being defended against.

CNBC — Amazon expects to spend $100 billion on capex in 2025 CNBC — Alphabet expects to invest about $75 billion in capex in 2025

WED · Feb 5, 2025modelsgooglepricingcontext windowagents

Gemini 2.0 Flash goes GA, Flash-Lite arrives, 2.0 Pro opens up

Google made Gemini 2.0 generally usable: 2.0 Flash became generally available with higher rate limits and simplified single-rate pricing, 2.0 Flash-Lite entered public preview as the most cost-efficient Gemini yet, and 2.0 Pro landed as an experimental update aimed at coding and complex prompts.

The Flash tier ships a 1 million token context window, native tool use and multimodal input, with image and audio output promised. Google claims performance gains across benchmarks over Gemini 1.5 while cutting effective cost on mixed-context workloads — and the same day GitHub made 2.0 Flash available inside Copilot Chat.

WHY IT MATTERS

Million-token context with native tool calling at a cut rate is the cheapest place to run long-horizon agent loops this week. Flash-Lite also resets the price anchor for high-volume, low-reasoning agent subtasks like extraction, routing and summarization — the tier that decides whether a multi-agent fan-out pencils out.

Google Developers Blog — Gemini 2.0: Flash, Flash-Lite and Pro Google — Gemini 2.0 is now available to everyone

TUE · Feb 4, 2025regulationeu ai actsafetyalignmentjailbreaks

EU publishes its prohibited-AI rules the same day Google deletes its weapons and surveillance pledges

The European Commission published its guidelines on the AI Act's prohibited practices, covering harmful manipulation, social scoring and real-time remote biometric identification, two days after the prohibitions themselves became applicable — the Act's first compliance deadline, with fines reaching €35M or 7% of global turnover once enforcement powers bite. The guidelines are non-binding; authoritative interpretation sits with the Court of Justice.

On the same day, Google removed from its AI principles the 2018 pledges not to use the technology for weapons development or surveillance. Also this week in safety research, Anthropic published Constitutional Classifiers, a synthetic-data guard that held up through more than 3,000 hours of human red teaming with no successful universal jailbreak in its target CBRN domain.

WHY IT MATTERS

European regulators were writing the binding interpretation of what AI must never do while a US frontier lab deleted its own self-imposed red lines. For anyone shipping agents into European markets the prohibited-practice list is now the compliance baseline, and guard-classifier results like Anthropic's are the cheap control that keeps a jailbreak from becoming a compliance event.

European Commission — guidelines on prohibited AI practices CNBC — Google removes pledge to not use AI for weapons, surveillance Anthropic — Constitutional Classifiers

SUN · Feb 2, 2025agentsopenairesearch agentstool useevals

OpenAI launches the deep research agent, its first production autonomous worker

OpenAI released deep research in ChatGPT: an agent that plans, browses and synthesizes hundreds of sources over tens of minutes into a cited analyst-grade report, powered by a version of the upcoming o3 model optimized for web browsing and data analysis.

OpenAI ships it with full citation trails and a visible thinking summary. Rollout started with Pro, and was extended to Pro users in the UK, Switzerland and the EEA within days. Reporting on the launch benchmark put it around 26% on Humanity's Last Exam.

WHY IT MATTERS

This is the template for long-horizon research agents: budgeted browsing, citation fidelity, personal-data and prompt-injection exposure from arbitrary web pages, and per-query cost measured in tens of minutes of inference. Nobody shipped the harness, only the output — so the engineering lesson is that multi-hour, tool-using, self-pivoting search loops are now a consumer product tier.

OpenAI — Introducing deep research Fortune — OpenAI's deep research and Humanity's Last Exam

DISPATCH 05

WEEK OF JAN 25 – 31, 2025

A downloadable model erased $593B from Nvidia, then everyone pushed back

DeepSeek's app took the top of the App Store and wiped nearly $600 billion off Nvidia in a single session. Within 72 hours a privacy regulator, a competitor, a security firm and a rival CEO had all responded.

Nvidia closed down about 17% on Monday, shedding close to $593 billion in market value — the largest single-day market-cap loss ever for a US company — as investors repriced the assumption that frontier AI requires ever-larger GPU fleets. The trigger was a free assistant built on the openly downloadable R1 model.

The response arrived from every direction: Italy's Garante blocked the service on GDPR grounds, Wiz found an unauthenticated production database leaking chat logs and API keys, OpenAI claimed DeepSeek distilled its models, and Anthropic's CEO argued the moment made chip export controls more important, not less.

FRI · Jan 31, 2025openaireasoningmodelsagent toolingsafety

OpenAI answers DeepSeek with o3-mini — and the first model rated Medium on autonomy

OpenAI released o3-mini in ChatGPT and the API, its cheapest reasoning model so far and the first small reasoning model with function calling, Structured Outputs and developer messages built in — the pieces agent builders actually wire together.

Three reasoning-effort levels let callers trade latency for depth. OpenAI says it produced math and factuality answers about 24% faster than o1-mini with 39% fewer major mistakes, tripled the Plus and Team limit from 50 to 150 messages per day, and put a reasoning model in front of free-plan users for the first time. There is no vision support. The accompanying system card rates CBRN, Persuasion and Model Autonomy Medium and Cybersecurity Low.

OpenAI answers DeepSeek with o3-mini — and the first model rated Medium on autonomy — Mistral AI

WHY IT MATTERS

The contrast is the story: six days after a free, openly downloadable Chinese reasoner broke the assumption that frontier quality requires frontier spend, OpenAI's counter was cheaper closed reasoning plus a safety scorecard that admits a real autonomy shift. For anyone building harnesses, the function calling, structured outputs and reasoning-effort surface matters more than the benchmark table — it makes a reasoning model usable as a tool-calling component rather than a chatbot.

OpenAI — OpenAI o3-mini OpenAI — o3-mini System Card

THU · Jan 30, 2025open weightsmistralself-hostingefficiencyfunction calling

Mistral Small 3: a 24B Apache-2.0 model that runs on one RTX 4090 and matches Llama 3.3 70B

Mistral released Mistral Small 3, a latency-optimized 24B dense model under Apache 2.0 — both pretrained and instruction-tuned checkpoints, 32K context, native function calling and JSON output, with strong system-prompt adherence.

It scores above 81% on MMLU at roughly 150 tokens per second, and Mistral claims parity with Llama 3.3 70B at three times the parameter count while being more than 3x faster on the same hardware; quantized, it fits a single RTX 4090 or a 32GB MacBook. Notably it was trained without RL and without synthetic data, positioning it as a base for reasoning work rather than a finished reasoner. Mistral simultaneously recommitted to Apache 2.0 for general-purpose models, moving away from its restrictive MRL license.

Mistral Small 3: a 24B Apache-2.0 model that runs on one RTX 4090 and matches Llama 3.3 70B — Mistral AI

WHY IT MATTERS

A second frontier-adjacent open-weight release in the same week pushes the deployable cost floor down to one consumer GPU: agent-grade function calling with no per-token bill, no data leaving the machine, and no closed-lab terms of service. That is the practical version of the cheap-intelligence argument the market was repricing on Monday — and note the license change, which matters as much as the weights.

Mistral AI — Mistral Small 3 Hugging Face — Mistral-Small-24B-Instruct-2501

THU · Jan 30, 2025regulationexport controlsprivacydeepseekpolicy

Italy blocks DeepSeek; OpenAI alleges distillation; Amodei says controls matter more

Italy's data protection authority imposed an immediate and definitive limitation on DeepSeek's processing of Italian users' personal data after the two Chinese companies behind the app answered its questions in a way it called totally insufficient and argued the GDPR did not apply to them. The app had already disappeared from Italian app stores the previous day.

The same week, OpenAI told the FT it had evidence DeepSeek used its proprietary models to train R1 via distillation — a terms-of-service violation — and the White House AI and crypto czar said there was substantial evidence it had happened. On 29 January Anthropic CEO Dario Amodei published an essay arguing DeepSeek's cost figures were on-trend at best rather than a discontinuity, that R1 essentially replicates OpenAI's o1 RL approach, and that cheaper training makes chip controls more important because million-chip clusters cannot be smuggled the way thousands can.

WHY IT MATTERS

This is the week's explicit contrast: an openly downloadable model anyone can self-host, landing next to a privacy regulator shutting the service down, a competitor alleging its outputs were scraped, and a frontier-lab CEO arguing for harder compute controls. Note what the essay never mentions — that DeepSeek ships open weights at all, and that the weights are already out regardless of what happens to the chips.

Reuters — Italy's regulator blocks DeepSeek on data protection Dario Amodei — On DeepSeek and Export Controls The Verge — OpenAI has evidence its models helped train DeepSeek

WED · Jan 29, 2025securitydeepseekmisconfigurationinfrastructuredata exposure

Wiz finds DeepSeek's production database wide open: chat history and API keys, no authentication

Wiz Research, assessing DeepSeek's external security posture days after R1 went viral, found a ClickHouse database on DeepSeek subdomains reachable from the open internet with no authentication at all.

It allowed full database control and potential privilege escalation inside DeepSeek's environment, and held over a million lines of log streams including plaintext chat history, API keys, backend hostnames and operational metadata. Wiz disclosed responsibly and DeepSeek secured it the same day.

WHY IT MATTERS

The security story of the open-model rush is not model jailbreaks — it is the ordinary infrastructure bolted onto fast-growing AI products. The same failure mode, an unauthenticated internal service reachable from the internet holding credentials and user data, is exactly what kills self-hosted agent harnesses, MCP servers and coding agents. That is why it belongs next to the model news rather than in a separate infrastructure column.

Wiz Research — exposed DeepSeek database leaking chat history Computer Weekly — DeepSeek exposure a 'rookie' cyber error

TUE · Jan 28, 2025open weightsmultimodalagentscomputer usechinaalibaba

Alibaba answers: Qwen2.5-Max claims to beat DeepSeek-V3, and Qwen2.5-VL ships open weights with computer use

Alibaba announced Qwen2.5-Max, a large-scale MoE pretrained on more than 20 trillion tokens and post-trained with curated SFT and RLHF, available only through the Alibaba Cloud API — no downloadable weights. Qwen claims it outperforms DeepSeek-V3 on Arena-Hard, LiveBench, LiveCodeBench and GPQA-Diamond.

Two days earlier Qwen had open-sourced Qwen2.5-VL in 3B, 7B and 72B sizes, a vision-language model that acts as a visual agent able to direct tools and perform computer and phone use without task-specific fine-tuning, understand videos over an hour with event localization, emit bounding boxes and points, and produce stable JSON structured output. Qwen2.5-VL-7B is claimed to beat GPT-4o-mini on several tasks.

WHY IT MATTERS

China's second frontier lab proves R1 was not a one-off, and splits its strategy: the flagship stays behind an API while the agentic multimodal model goes out as open weights. Computer use, tool direction and stable JSON output are precisely the harness-shaped capabilities builders were paying closed labs for — here they are a download, and the split tells you where each lab thinks its durable advantage sits.

Qwen — Qwen2.5-Max Qwen — Qwen2.5-VL release

MON · Jan 27, 2025marketsnvidiaopen weightsinference costdeepseek

DeepSeek takes the App Store and erases ~$593B from Nvidia in one day

Nvidia closed down about 17% on Monday, shedding close to $593 billion in market value — the largest single-day market-cap loss ever for a US company — while AI-exposed semiconductor, power and infrastructure shares collectively dropped more than $1 trillion as investors repriced the assumption that frontier AI requires ever-larger GPU fleets.

The trigger was DeepSeek's free assistant, built on the openly downloadable R1 model and claimed to be trained for a fraction of incumbent cost. The app hit number one on the US App Store, up from number 31 days earlier, and was top free app in the US and 51 other countries, with downloads doubling from 1M to 2.6M in days. Trump called it a wakeup call; Altman called the model impressive. Nvidia rebounded 8.9% the next day.

WHY IT MATTERS

The clearest evidence yet that open weights can move the economics of compute: a freely downloadable model, not an API product, forced a trillion-dollar markdown in a single session. It sets up everything else this week — cheaper closed reasoners from OpenAI, permissive open releases from Mistral, and a policy fight about whether the chips that produced it should have been exported at all.

Reuters — DeepSeek sparks AI stock selloff CNBC — Nvidia sheds almost $600 billion in market cap, biggest drop ever

DISPATCH 04

WEEK OF JAN 18 – 24, 2025

DeepSeek gave away o1-class reasoning — the same week Washington committed $500B and scrapped its safety order

An MIT-licensed 671B model landed on Monday. By Thursday the US had announced half a trillion dollars of compute and deleted the frontier-safety review regime it had created fifteen months earlier.

DeepSeek-R1 shipped as open weights under the MIT license — 671B total parameters, 37B active, 128K context — claimed on par with OpenAI's o1 on math, code and reasoning at a fraction of the API price, alongside six distilled dense models from 1.5B to 70B.

Within 72 hours: Stargate was announced at the White House with $100B committed immediately, and Trump signed EO 14179 revoking Biden's AI safety order. OpenAI also shipped Operator, a cloud browser your agent drives, with prompt-injection mitigations security researchers immediately doubted.

FRI · Jan 24, 2025computemetacapexdata centers

Meta sets 2025 capex at $60–65B, targeting 1.3M GPUs

Zuckerberg announced that Meta will spend $60–65 billion on capital expenditures in 2025, up from $38–40B in 2024 and well above the roughly $51B Wall Street consensus, building a 2GW+ data center, bringing about 1GW of compute online this year and ending 2025 with more than 1.3 million GPUs.

He called 2025 a defining year for AI, predicted Llama 4 becomes the leading state-of-the-art model, said Meta AI should serve a billion-plus people, and said Meta is building an AI engineer that will start contributing increasing amounts of code to its R&D efforts. Meta shares closed at a record.

Meta sets 2025 capex at $60–65B, targeting 1.3M GPUs — image.cnbcfm.com

WHY IT MATTERS

The compute side of the week's contrast: closed labs answering an open-weight cost shock with more silicon, not cheaper models. It is also a bigger number than the entire 2024 frontier-training spend of most labs, placed days after R1 showed o1-class reasoning under an MIT license.

CNBC — Zuckerberg sets Meta's AI targets for the year Reuters — Meta to spend up to $65 billion this year to power AI goals

THU · Jan 23, 2025agentsagent harnessprompt injectionbenchmarksopenai

OpenAI ships Operator and the Computer-Using Agent: a cloud browser your agent drives

OpenAI released Operator as a research preview for ChatGPT Pro users in the US — a cloud-hosted Chrome instance the model drives by screenshot, typing, clicking and scrolling, with no site APIs involved. It is powered by CUA, which pairs GPT-4o vision with RL-trained GUI reasoning.

It set state of the art on computer and browser-use benchmarks: 38.1% on OSWorld against a prior 22.0% and a 72.4% human baseline, 58.1% on WebArena, and 87% on WebVoyager. The harness-level safety design is the interesting part: cautious navigation against prompt injection, a separate monitor model that ingests the screenshot and pauses execution on suspicious content, takeover mode so passwords and card details never enter the model context, restricted site categories, and a system card rating autonomy risk as Low.

OpenAI ships Operator and the Computer-Using Agent: a cloud browser your agent drives — static.simonwillison.net

WHY IT MATTERS

The reference harness design for the whole browser-agent category this year — treat its eval numbers and its prompt-injection posture as the baseline your own agent stack gets compared to. Simon Willison's reaction was the fair counterpoint: indirect prompt injection from a hostile page is exactly the failure mode that broke Claude Computer Use in 2024, and the mitigations are ML classifiers, not a sandbox boundary.

OpenAI — Introducing Operator OpenAI — Computer-Using Agent research Simon Willison — Introducing Operator (security analysis)

THU · Jan 23, 2025regulationpolicyexecutive ordersafety

Trump signs EO 14179, revoking Biden's AI safety order

Executive Order 14179, Removing Barriers to American Leadership in Artificial Intelligence, declares it US policy to sustain and enhance America's global AI dominance, revokes Biden's EO 14110 of 30 October 2023 — the Safe, Secure and Trustworthy AI order that created reporting duties for frontier models — and orders agencies to identify and suspend, revise or rescind actions taken under it.

It also instructs an AI Action Plan within 180 days and directs OMB to revise its AI memoranda within 60 days.

WHY IT MATTERS

The dated primary source for where US frontier-AI policy reset — no reporting or preemption frameworks survive it, and everything downstream this year cites it. It arrived within 72 hours of an unaffiliated Chinese lab publishing MIT-licensed weights matching o1-class reasoning, and two days after the White House announced $500B of private AI infrastructure.

The White House — Removing Barriers to American Leadership in AI GovInfo — Executive Order 14179, full text

WED · Jan 22, 2025fundingindustryanthropicgooglevaluation

Google adds another $1B to Anthropic

Google agreed to invest more than $1 billion more in Anthropic, on top of the roughly $2B it had already put in and its stake of about 10%, confirmed by CNBC after an FT report.

It lands alongside Anthropic's late-stage talks to raise about $2B at a $60B valuation led by Lightspeed, and after Amazon's additional $4B in November 2024. CNBC's source said Anthropic's revenue hit an approximately $1B annualized run-rate in December, roughly ten times year over year, mostly enterprise.

WHY IT MATTERS

Two hyperscalers now fund the same safety-focused lab, and the valuation jump is the clearest read on what enterprise demand was priced at going into 2025. The defensive move against the week's open-weight shock is to buy equity in a rival lab rather than to open-weight your own models.

CNBC — Google agrees to new $1 billion investment in Anthropic Reuters

TUE · Jan 21, 2025computedata centersstargateopenaisoftbankcapex

Stargate announced at the White House: $100B now, up to $500B over four years

Trump announced Stargate in the Roosevelt Room with Sam Altman, Larry Ellison and Masayoshi Son: a new company intending to invest $500 billion over four years in AI infrastructure in the US, starting with $100B deployed immediately. Initial equity funders are SoftBank, OpenAI, Oracle and MGX, with SoftBank holding financial responsibility and OpenAI operational responsibility.

Buildout starts in Texas; Ellison said 20 data centers of roughly 500,000 square feet each are planned. In parallel Microsoft reset its OpenAI terms: a new large Azure commitment and IP and revenue-share continuity through 2030, but exclusivity on new capacity replaced by a right of first refusal.

WHY IT MATTERS

The compute-commitment side of the week, and a structural change that mattered more than the headline number — losing capacity exclusivity let OpenAI buy compute from anyone. Note the sequencing: this half-trillion-dollar commitment was announced the day after an MIT-licensed model from Hangzhou matched o1 on reasoning benchmarks, which is precisely the argument DeepSeek's release undercut.

OpenAI — Announcing The Stargate Project Reuters — Trump announces private-sector $500 billion investment in AI infrastructure

MON · Jan 20, 2025open weightsdeepseekreasoningrlself-hostingdistillation

DeepSeek-R1: MIT-licensed open weights at o1-class reasoning

DeepSeek released R1 as open weights on Hugging Face under the MIT license — 671B total parameters, about 37B active per token, 128K context — plus R1-Zero, which is pure reinforcement learning with no supervised cold start, and six distilled dense models from 1.5B to 70B built on Qwen and Llama checkpoints.

The model card claims performance comparable to OpenAI's o1 across math, code and reasoning, with AIME 2024 at 79.8 and MATH-500 at 97.3. DeepSeek-R1-Distill-Qwen-32B is claimed to beat o1-mini, and coverage put API pricing roughly 90 to 95% below o1. The paper's key claim is that reasoning behavior emerged from large-scale RL without human-labeled reasoning trajectories.

WHY IT MATTERS

The event the rest of the week is a response to — the point where an MIT-licensed model became the price-setting alternative to closed frontier reasoning. Commercial use, self-hosting, fine-tuning and distillation were all permitted, which sets a hard cost floor under closed reasoning APIs, makes a 37B-active model self-hostable on serious single-node hardware, and hands every lab and enterprise a distillation source.

Hugging Face — deepseek-ai/DeepSeek-R1 (MIT license)arXiv — DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via RL

DISPATCH 03

WEEK OF JAN 11 – 17, 2025

Washington moved to license model weights. Two days later a 456B model shipped openly.

The US extended export controls to AI model weights for the first time — and a Chinese lab answered by putting frontier-adjacent long-context weights on Hugging Face with MIT-licensed inference code.

On Monday the outgoing US administration published the AI Diffusion framework, extending worldwide licensing to advanced AI chips and, for the first time, to the weights of closed frontier models. Nvidia called it a 200-plus-page overreach that would hand the global AI market to Chinese competitors.

Two days later MiniMax pushed the weights of a 456-billion-parameter model with a 4-million-token context to Hugging Face. Also this week: Snyk published npm packages named after Cursor's internal extensions while testing dependency confusion, and Apple switched off AI summaries of news notifications after the BBC complained the feature was inventing headlines.

THU · Jan 16, 2025deploymentsafetyapplehallucinationmedia

Apple disables AI news summaries after they invented headlines

Apple switched off Apple Intelligence notification summaries for the News and Entertainment category in the iOS 18.3, iPadOS 18.3 and macOS Sequoia 15.3 betas, saying the feature would be temporarily unavailable while it works on improvements.

The summaries were optional and limited to the newest iPhones, but they carried the news organization's own branding, and they were wrong repeatedly: one alert told readers the man accused of killing UnitedHealthcare CEO Brian Thompson had shot himself. The BBC complained in December 2024 and got no reply until January; Reporters Without Borders and the National Union of Journalists had both called for withdrawal, and Apple's initial fix of italicizing the text to mark it as AI-generated satisfied nobody.

Apple disables AI news summaries after they invented headlines — ichef.bbci.co.uk

WHY IT MATTERS

The clearest consumer-facing example of a summarization product damaging real publishers: the failure was not a stale index or a bad ranking, it was a model writing sentences no source ever contained and attaching them to a trusted masthead. For anyone shipping summarization over untrusted documents — agent receipts, PR digests, log triage — attribution and confidence are the product, not fluency.

BBC News — Apple suspends error-strewn AI generated news alerts The Washington Post — Apple pauses AI summaries that botched news headlines

WED · Jan 15, 2025open weightsminimaxmixture of expertslinear attentionlong contextagents

MiniMax open-sources a 456B-parameter model with a 4M-token context

MiniMax released the MiniMax-01 series — MiniMax-Text-01 and the vision-language MiniMax-VL-01 — as downloadable weights on Hugging Face and GitHub, with a 68-page technical report.

It is a 456 billion parameter mixture-of-experts model with 45.9B active parameters per token, 80 layers and 32 experts under top-2 routing, and it replaces most attention with Lightning Attention: one softmax attention layer after every seven linear ones, which is what lets it train at a 1M-token context and extrapolate to 4M at inference. On RULER it scores 0.910 at 1M tokens where Gemini 1.5 Pro manages 0.850, and on LongBench v2 it posts 56.5 overall against GPT-4o's 51.4 — above the 53.7 human baseline. MiniMax reports 100% retrieval on a 4M-token needle test. The weights are under a custom MiniMax Model Agreement while the inference code is MIT, and API pricing is $0.20 per million input tokens.

MiniMax open-sources a 456B-parameter model with a 4M-token context — filecdn.minimax.chat

WHY IT MATTERS

An openly downloadable model that is frontier-adjacent specifically on the axis that matters most for agents — long context — at a price that makes self-hosting and cheap fan-out plausible. The architectural bet, that you can trade softmax attention for linear attention at 456B scale without falling off a cliff, attacks inference cost rather than benchmark scores, which is where the closed labs' margin lives.

MiniMax — MiniMax-01 is now open source Hugging Face — MiniMaxAI/MiniMax-Text-01 VentureBeat — MiniMax unveils its own open-source LLM with a 4M token context

TUE · Jan 14, 2025securitysupply chaincoding agentsnpmcursor

Snyk published npm packages named after Cursor's internals while testing dependency confusion

Three packages appeared on npm — cursor-retrieval, cursor-always-local and cursor-shadow-workspace — published by a user under a verified snyk.io email address. They were named after extensions bundled inside Cursor, the AI code editor, and on install they phoned home with the username, hostname, working directory and environment variables to a researcher-controlled server. OpenSSF's package analysis scanner flagged all three as malicious.

Snyk's CTO confirmed it was Snyk Security Labs testing whether Cursor's build pulled VSCode-style extensions from the public registry, that the packages were labeled as test packages and time-boxed to 24 hours, and that the team found no indication Cursor was vulnerable. Anysphere's co-founder said Cursor had not hired Snyk, that Snyk apologized, and that the theory it was an npm error could not hold because the names were Cursor's bundled extensions, which are not published to a registry at all.

WHY IT MATTERS

The argument was never really about Snyk. It is that an AI coding tool's build and extension chain is a public namespace anyone can squat, and that a package which merely reads environment variables is enough to hand over GitHub tokens, cloud keys and npm credentials — the exact credentials a coding agent runs with.

The Register — Snyk deployed 'malicious' packages, claims infoseccer Snyk — Security Labs testing update: Cursor.com AI code editor

TUE · Jan 14, 2025computedata centersenergypolicyexecutive order

Biden orders gigawatt AI data centers onto federal land

Executive Order 14141 directs the Departments of Defense and Energy to make federal sites available to be leased for frontier AI data centers and the generation to power them, with each department to identify at least three candidate sites by 28 February 2025 and the Interior Department to nominate suitable public lands by 15 March — aiming at sites fully permitted by the end of 2025.

Developers selected will be contractually required to bring online clean generation matching the full electricity demand of their data centers, plus safeguards around national security and community impact. Agencies are also ordered to study the effect of AI data centers on electricity prices.

WHY IT MATTERS

The compute build-out is now a land-and-power question, and this is the US government volunteering federal acreage and permitting priority to shorten it. Note what is actually scarce: not land, but interconnection and clean generation — which is why the order makes matched clean power a contractual term.

The White House — Executive Order on AI infrastructure GovInfo — Executive Order 14141 (PDF)

MON · Jan 13, 2025export controlsregulationmodel weightschipsnvidiaopen weights

US extends export controls to AI model weights — two days before a 456B Chinese model ships openly

The Bureau of Industry and Security published the Framework for Artificial Intelligence Diffusion, effective immediately with compliance from 15 May 2025. It puts a worldwide license requirement on advanced computing ICs under ECCNs 3A090.a and 4A090.a and creates ECCN 4E091, a new worldwide licensing requirement on the weights of closed-weight dual-use AI models trained above 10^26 operations — the first time US export law has reached model weights rather than hardware.

Chip flows are governed by per-country total-processing-performance quotas, with 18 listed allies treated favorably and a presumption of denial for Tier 3 destinations. A companion interim final rule extended licensing and due-diligence duties to foundries and packaging firms. Nvidia's government affairs lead called it unprecedented and misguided, a 200-plus-page regulatory morass drafted in secret, and warned that we are likely handing most of the global AI and GPU market to our Chinese competitors.

WHY IT MATTERS

Put the two halves of this week side by side. Washington spent Monday drafting a license regime for the weights of closed models and quotas for the chips that train them; on Wednesday MiniMax put 456 billion frontier-adjacent parameters under a downloadable file with MIT-licensed inference code, at a price no quota system can police once the weights are on the internet. Nvidia's own objection is the sharpest thing anyone said about the rule, and it came from the company that sells the hardware.

Federal Register — Framework for Artificial Intelligence Diffusion NVIDIA — Statement on the AI Diffusion rule Hogan Lovells — analysis of the model-weight controls

MON · Jan 13, 2025industrymicrosoftleadershipreorganizationagentsplatform

Microsoft folds dev tools, AI platform and agent runtimes into one CoreAI division

Satya Nadella told staff that Microsoft is creating CoreAI — Platform and Tools, an engineering organization led by Jay Parikh, the former global head of engineering at Meta, as executive vice president reporting directly to Nadella.

It merges the full Developer Division, the AI Platform group, and a set of Office-of-the-CTO teams including AI Supercomputer and AI Agentic Runtimes. The stated mission is to build the end-to-end Copilot and AI stack for first-party and third-party customers to build and run AI apps and agents. Nadella's framing: agents will re-form every SaaS category, and 30 years of change is being compressed into three.

WHY IT MATTERS

A structural signal about where Microsoft thinks the value sits: not in owning a frontier model, but in owning the agent runtime and the tooling developers build agents with. Pulling AI Agentic Runtimes out of the CTO's office and into the same org as the developer division is a bet that the harness, sandbox and tool-calling layer is the durable platform, and that model access will be a commodity.

Microsoft — Introducing CoreAI CNBC — Nadella forms new CoreAI group led by Jay Parikh

Go back in time

Every dispatch, newest first. Each week is written once and left as it was published.

2026

SEPTEMBER

AUGUST

JULY