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 02

WEEK OF JAN 4 – 10, 2025

A $450 open model matched o1-preview the week Anthropic hit a $60B valuation

The first full working week of 2025 split the industry in two directions at once: open-weight recipes collapsing the cost of reasoning, and capital compounding at the closed labs.

UC Berkeley's NovaSky team released Sky-T1-32B-Preview, a fully open reasoning model fine-tuned in 19 hours on 8 H100s for under $450, with weights, data and code all published. Microsoft's Phi-4 shipped its full 14B weights under MIT.

Against that, Anthropic was reported in late-stage talks to raise up to $2B at a $60B valuation, and the FTC and DOJ filed a statement of interest in Musk v. Altman arguing that OpenAI and Microsoft board interlocks can breach the antitrust laws.

FRI · Jan 10, 2025open weightsreasoningpost-trainingdistillationcost

UC Berkeley's NovaSky open-sources Sky-T1-32B: o1-preview-class reasoning for under $450 of H100 time

The NovaSky team at UC Berkeley released Sky-T1-32B-Preview on Hugging Face, the first reasoning model it describes as fully open: weights, the 17K-sample training set, the data-generation and evaluation code, and a technical report are all downloadable.

Sky-T1 is a supervised fine-tune of the open Qwen2.5-32B-Instruct base on 17K curated traces generated by QwQ-32B-Preview with rejection sampling — 5K coding from APPs and TACO, 10K math from AIME, MATH and Olympiad subsets, 1K science and puzzle data — trained for 3 epochs in 19 hours on 8 H100s for roughly $450 at Lambda Cloud pricing. It reports Math500 82.4 against o1-preview's 81.4, AIME2024 43.3 against 40.0, and LiveCodeBench-Medium 56.8 against 54.9.

UC Berkeley's NovaSky open-sources Sky-T1-32B: o1-preview-class reasoning for under $450 of H100 time
Hugging Face

WHY IT MATTERS

A fully reproducible, sub-$500 recipe for o1-preview-class reasoning with all data and code released is the clearest statement yet that reasoning capability is not scarce. The fine-tuning floor collapsed: frontier-adjacent reasoning is now a single-node, single-day, few-hundred-dollar job that any lab, startup or hobbyist can copy and improve on, which undercuts the claim that reasoning justifies the closed-lab price premium.

FRI · Jan 10, 2025antitrustlitigationregulationopenaimicrosoft

FTC and DOJ argue OpenAI/Microsoft board interlocks can break antitrust law

The antitrust agencies filed a statement of interest in Musk v. Altman (N.D. Cal., Case 4:24-cv-04722-YGR), the private suit in which Elon Musk alleges OpenAI and Microsoft illegally monopolized the AI market.

The government did not take a position on who should win. It argued that a single person serving on the boards of two competing companies can violate the antitrust laws even if they forgo one of the seats, and that the agencies read interlocks broadly enough to capture board observers and arrangements that survive nominal resignation. At issue in the filing is Reid Hoffman's simultaneous participation on the OpenAI and Microsoft boards.

FTC and DOJ argue OpenAI/Microsoft board interlocks can break antitrust law
www.law360.co.uk

WHY IT MATTERS

The first time US antitrust enforcers have staked out an interlocking-directorate interpretation aimed squarely at the AI industry, filed in the case that also tags Microsoft's OpenAI stake. Applied literally it is uncomfortable for the whole sector: board overlap between hyperscalers and the model labs they fund is structural, and Microsoft held a non-voting OpenAI board observer seat after the November 2023 board crisis.

WED · Jan 8, 2025open weightssmall modelssynthetic datalicensingmicrosoft

Microsoft Research puts Phi-4's full weights on Hugging Face under an MIT license

Microsoft Research released the complete weights of Phi-4 to Hugging Face under an MIT license. Phi-4 is a 14B-parameter dense decoder-only model with a 16K-token context window, trained on 9.8T tokens over 21 days on 1,920 H100-80Gs, and it had been available before that only through Azure AI Foundry.

The training recipe is the point: the technical report says Phi-4 is built around synthetic data throughout pre-training and substantially surpasses its teacher model, GPT-4, on STEM-focused question answering — evidence that small, data-curated models beat scale-for-scale assumptions.

WHY IT MATTERS

A permissively licensed 14B model with published training detail, downloadable by anyone, that fits on a single consumer-class GPU — the free tier of frontier-adjacent reasoning got its most usable artefact of the week, and it removes the last license excuse for anyone still renting a larger closed model for STEM-heavy reasoning.

TUE · Jan 7, 2025fundingvaluationanthropicbusiness modelclosed weights

Anthropic in late-stage talks to raise up to $2B at a $60B valuation

CNBC reported that Anthropic was in late-stage talks to raise as much as $2 billion at a $60 billion valuation, with Lightspeed Venture Partners leading, confirming earlier Wall Street Journal reporting.

It is a more than threefold re-rating in about a year for a company that sells closed frontier models on a safety-first pitch, and it lands in the same five days as Sky-T1's $450 reasoning fine-tune and Phi-4's MIT-licensed weights.

WHY IT MATTERS

The contrast is the week's central tension: capital is compounding faster at the closed labs while the marginal cost of a frontier-adjacent reasoning model has already fallen to the price of a used car. The round reads as a bet on distribution, enterprise contracts and capital-intensive frontier training rather than on capability being scarce.

TUE · Jan 7, 2025regulationdeepfakesukharmscriminal law

UK moves to criminalize creating sexually explicit deepfakes

The UK Ministry of Justice published plans to make creating sexually explicit deepfake images a criminal offence, not only sharing them, with the offences to be carried in the Crime and Policing Bill.

The package also creates offences for taking an intimate image without consent and for installing or adapting equipment with intent to enable it, each carrying up to two years' custody. The deepfake offences cover images of adults, with children already protected by existing law. Ministers warned that platforms hosting the content will face tougher scrutiny and significant penalties.

WHY IT MATTERS

Concrete criminal liability, dated precisely, for creating non-consensual explicit synthetic media — a rare case of legislation moving faster than the model-release cycle, and it arrived while image and video generation capacity was being handed out openly and cheaply.

MON · Jan 6, 2025chipshardwarecesagent toolingopen weightsworld models

Nvidia opens CES with RTX 50-series GPUs, a $3,000 desktop AI computer and open Cosmos world models

Nvidia's CES keynote covered the full stack a model shop needs. Project DIGITS is a Grace Blackwell GB10 desktop system with 128GB of unified coherent memory that runs up to 200B-parameter models locally — two units linked for 405B — priced around $3,000, built to prototype and fine-tune locally then move to datacenter capacity. The Blackwell RTX 50 series arrived with the RTX 5090 at $1,999 and RTX 5080 at $999.

For builders, the Cosmos world foundation model platform shipped its first wave openly under Nvidia's open model license on Hugging Face and NGC for robotics and autonomous-vehicle synthetic data, alongside the announced Llama Nemotron and Cosmos Nemotron families for enterprise agents and the AI Blueprints built around them.

WHY IT MATTERS

Simultaneous movement on all three layers that matter: local inference hardware, agent-oriented model families, and openly licensed world models for robotics. One vendor now supplies the open weights, the agent models and the silicon they run on, which is why local inference of large models became a product category in this week rather than a demo.

DISPATCH 01

WEEK OF DEC 28, 2024 – JAN 3, 2025

A frontier-adjacent open-weight model for the price of a house meets $80B of capex

DeepSeek put a 671B flagship on Hugging Face for roughly $5.6M of training compute, and within days Microsoft answered with the largest datacenter number anyone had put on the record.

The cost floor of frontier-adjacent AI fell into public view this week. DeepSeek-V3 shipped with downloadable weights and a training-run figure in the single-digit millions of dollars.

Microsoft pledged roughly $80B of AI datacenter spending for fiscal 2025, more than half of it in the US. Meanwhile the US outbound-investment screening regime for Chinese chips, quantum and AI took legal effect, and the Kremlin instructed Russia's government to build AI cooperation with China.

FRI · Jan 3, 2025computedata centerscapexindustry

Microsoft commits ~$80B to AI datacenters in fiscal 2025

Brad Smith's Microsoft On the Issues post states the company is on track to invest approximately $80 billion in AI-enabled datacenters in fiscal 2025, with more than half of that total in the United States, framed as a pitch for the incoming administration to back American AI infrastructure, skilling and exports to allies.

Reuters and CNN both picked up the figure as a headline number for the sector's capex race. It landed the day after a Chinese lab published a frontier-adjacent open-weight model trained for a figure in the single-digit millions — the contrast that defined the week.

Microsoft commits ~$80B to AI datacenters in fiscal 2025
Microsoft

WHY IT MATTERS

The clearest public split in AI economics: a hyperscaler's infrastructure number set against an open-weight release that resets what a competitive model costs to train and to self-host.

THU · Jan 2, 2025litigationprivacyvoice assistantsapple

Apple agrees to pay $95M to settle Siri recording claims

Apple agreed to pay $95 million in cash to settle a proposed class action alleging its Siri assistant violated users' privacy by recording private conversations after accidental activations and sharing them with third parties.

The preliminary settlement was filed in Oakland federal court and needs approval from US District Judge Jeffrey White. The class period runs from 17 September 2014 to 31 December 2024, and members may receive up to $20 per Siri-enabled device, with the class estimated in the tens of millions. Apple denied wrongdoing. A parallel suit over Google's Voice Assistant, brought by the same plaintiffs' firms, is pending in the same district.

Apple agrees to pay $95M to settle Siri recording claims
www.federalregister.gov

WHY IT MATTERS

Voice-assistant data collection is the consumer-side version of the data-provenance fight the whole industry is in — and the settlement lands on the last day of the class period it covers.

THU · Jan 2, 2025regulationexport controlsinvestmentchina

US outbound-investment screening on Chinese AI, chips and quantum takes legal effect

The Treasury rule implementing Executive Order 14105 — 31 CFR Part 850, published at 89 FR 90398 — became effective on 2 January 2025. It bars or requires notification for certain US-person investments in Chinese, Hong Kong and Macau companies working on semiconductors and microelectronics, quantum information technologies, and artificial intelligence, the three sectors the order designated as national-security technologies.

Law-firm analyses of the regime circulated that week as the compliance date hit.

WHY IT MATTERS

For anyone building or funding AI with a China nexus, this is the operative compliance date: it turns export-control logic inward onto capital flows, and it sits in direct tension with the same week's open-weight release out of Hangzhou.

WED · Jan 1, 2025geopoliticschinarussiacompute

Putin orders Russian government and Sberbank to build AI cooperation with China

President Vladimir Putin ordered Russia's government and its largest bank, Sberbank, to build cooperation with China in artificial intelligence.

The instructions were published on the Kremlin website as sanctions pressure continues to limit Russian access to Western AI hardware and models.

WHY IT MATTERS

A state-directed alternative bloc is forming around compute and models, right as US investment restrictions on Chinese AI took effect the next day — a signal about where open-weight distributions travel when export controls bite.

MON · Dec 30, 2024computem&aantitrustgpu schedulingopen source

Nvidia closes its $700M Run:ai acquisition and plans to open-source the software

Nvidia completed its acquisition of Israeli AI infrastructure startup Run:ai after the European Commission granted unconditional approval earlier in December. The Commission had initially said the deal would require EU clearance and warned it could threaten competition in GPU-related markets.

Run:ai said it plans to make its GPU-virtualization and scheduling software open source so it can extend beyond Nvidia GPUs to the wider AI ecosystem. Nvidia holds roughly 80% of the AI GPU market.

WHY IT MATTERS

Two developer-relevant facts in one: Nvidia is buying the layer that schedules GPU capacity for AI workloads — what your training jobs queue on — and it intends to open-source it, which is a distribution move for the CUDA moat at the same moment regulators are watching the deal.

SAT · Dec 28, 2024open weightsMoEmodel releaseinference costchina

DeepSeek-V3: a 671B open-weight MoE reaches closed-frontier quality for a single-digit-million training run

DeepSeek released DeepSeek-V3 with downloadable weights, and the analysis ran through the New Year week. The specification: 671B total parameters with 37B activated per token, a 128K context window, pre-trained on 14.8 trillion tokens using multi-head latent attention, DeepSeekMoE and an auxiliary-loss-free load-balancing strategy.

The report's own framing is that it outperforms other open-source models and achieves performance comparable to leading closed-source models, and that the final training run needed only 2.788M H800 GPU-hours — the number behind the widely cited ~$5.6M compute estimate. Code is MIT-licensed; the weights ship under DeepSeek's own model license.

WHY IT MATTERS

This is the open-weight story of the quarter and the direct counterweight to the same week's $80B capex pledge and new China investment bans: frontier-adjacent capability that anyone can download and self-host, at an API price floor the closed labs have to answer, with 37B active parameters setting the realistic self-hosting bar.

ARCHIVE

Go back in time

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