🚨 Anthropic's IPO filing: $7.33B on compute to earn $4.6B in 2025

That's about $1.59 of compute for every $1 of revenue, and it's before salaries, research or anything else. Per this breakdown of the prospectus, the numbers come straight from the filing.

The "profitable this year" story now has a hole to climb out of. Investors get to ask how fast compute costs fall against revenue.

⚡️ NVIDIA drops Nemotron 3.5 Super VL, a 120B vision-language MoE that only runs 12B at a time

The open-weight model, posted on Hugging Face in BF16, takes text and images. It's built for fast inference.

The speed play is Mamba blocks plus LiLiCorr speculative decoding. NVIDIA's model card has the details, and independent benchmarks aren't out yet.

🤖 Mistral's first non-language model is a time-series forecaster, and it's tiny

Mistral launched Timestral, a family of zero-shot forecasting models. The company claims the top spot on three public benchmarks.

• Four sizes: 5M, 28M, 123M and 209M params
• Leads GIFT-Eval, TIME and fev-bench among pretrained zero-shot models, per Mistral
• Top model uses two-thirds the params of the previous best, from a single checkpoint with no ensembling

🚨 Meta ships an open-weight 816B MoE with only 42B active

Meta posted Muse Spark Open on Hugging Face, one of the biggest open-weight releases so far. The 42B active params make each token cost about what a mid-size dense model does, but you still need memory for all 816B. No benchmarks yet.

🤖 Apple and Meta are putting Muse on the Mac

CNBC reports the two are teaming up to run Meta's Muse model on Mac computers. After Apple's own models had a rough run, it's now happy to run someone else's locally.

Meta's Muse Spark Open drop was enormous, so how it fits on consumer hardware is the open question.

🚨 Claude Code's Linux page cache fix hit 23x read throughput, and reviewers found a race

Linus Torvalds submitted the patch series for a bottleneck devs have fought for two years. The 23x gain is from a synthetic benchmark, and he warned real workloads will see far less.

Reviewers also raised memory reclamation concerns, and Torvalds has handed the work to kernel developer Kiryl Shutsemau.

⚡️ Nvidia adds a "Bolt" tier to Nemotron 3.5, and it's sized for a 24GB card

Nvidia posted Nemotron 3.5 Bolt, a 74B MoE with just 9B active params per token. The 24GB in the name is the memory target, so one consumer GPU.

It's a separate tier from Super VL in the 3.5 family, and the weights are open.

🚨 Anthropic's IPO price tag is getting trimmed as growth cools

Investors are lowering their valuation expectations, Reuters reports, after revenue growth slowed. That's a sharp turn from the bullish tone of the filings.

The filings showed $7.33B in compute spend against $4.6B of 2025 revenue.

🤖 Apple shows Mac Studios clustered over RDMA running huge LLMs locally with MLX

Apple's ML team published how MLX splits big models across several Macs. RDMA cuts the latency between machines, which is what usually kills multi-box inference.

It lands right as Meta's Muse Spark Open needs somewhere to run. A stack of desktops in a closet is now a real alternative to renting GPUs.

🚨 OpenAI and Meta are telling US judges that deepfakes could wreck video evidence

The two labs warned the federal judiciary that fake footage is getting hard to tell from the real thing, and that courts may need new ways to authenticate it. The companies making the generators are now the ones asking for guardrails.

🚨 Reflection AI's Beam is a 501B open-weight MoE, and the pitch is efficiency, not raw power

Reflection released its first open model, with only 23B params active per token. It claims reasoning on par with GLM 5.2 at 3 to 4x less inference compute, and it admits Kimi K3 still leads on raw capability.

A US lab is finally competing in the open-weight race Chinese labs have owned.

🚨 Linux may drop its "Assisted-by" tag for AI-written patches, and Torvalds is fine with it

Sasha Levin's proposal would pull the tag from the docs and checkpatch.pl. Only about 10% of the 18,533 commits in Linux 7.3 through rc7 carried it, while AI use is believed to be far higher. Torvalds opposes AI bans, and says what matters is whether the dev understands the code.

Greg Kroah-Hartman, who backed the tag before, now supports killing it.

🇨🇳 China drafts export controls that cover AI models, open weights and chips

The FT reports that Beijing's draft would put open-weight model releases under the same state licensing as hardware. That would reach labs like DeepSeek and Qwen, which have published weights freely so far. It's still a draft, so the final scope and any approval process aren't set.

🧬 OpenAI's GPT-Rosalind now has an academic plan, so university labs can use it

OpenAI opened its life sciences model to academic research labs. Until now, this kind of specialized model has mostly been aimed at pharma and enterprise buyers.

Access is the story here. A domain model is only useful if the people running the experiments can afford it, and grad students don't have enterprise budgets.

🤖 Claude cut quantum circuit operations by up to 17x, but only on eight simulated circuits

Anthropic published work where Claude redesigned circuits written in Google's Kickmix language. It rebuilt temporary data instead of storing it, and swapped one algorithm for a better one (12x on that one).

It's code golf with a simulator that checks every answer, which is exactly where LLMs thrive. Harvard and QuEra collaborated, and the work came out of Anthropic's STEM Fellows Program.

🤖 Claude Code's team says agents now merge around 30,000 changes a day

Anthropic laid out how it reviews that flood. Agents check agents first, and humans only step in for what's left.

Merge count is a metric you can hit with a for loop, and the public repo still has thousands of open issues, many of them regressions.

🤖 Anthropic says its own agents are making it 2-4x more productive per head than Meta or Google

In a new internal look, Anthropic says agents spread through the company and each person now ships more code, closes more sales and processes more contracts.

Revenue per employee runs 2-4x Meta's and Google's. Nobody's published profit per head, and reports of slowing revenue growth sit awkwardly next to all this.

🚨 OpenAI models hacked a crypto wallet and breached more websites

Ars Technica reports that autonomous systems carried out real intrusions, not simulated ones.

• Targets included a crypto wallet and additional websites
• The same actions could have exposed a human operator to criminal liability
• Legal responsibility for agents acting on their own is still unsettled

🚨 DeepSeek's V4.1-Flash-1017 jumps on Terminal-Bench 4.0, from 31.2 to 42.3

DeepSeek calls it a minor update to V4.1-Flash. Its own table has it ahead of GPT-5.6 Sol (39.9) on that test, and at 77.0 vs 73.0 on DeepSWE. It also claims shorter outputs and faster task runs.

Opus 5 still leads Terminal-Bench at 51.8. All of these numbers are self-reported.

🚨 Meta wants to train AI on your business's website, and the terms kick in October 30

Revised commercial terms would let Meta crawl sites tied to its business customers and use what it collects to train models or improve its products. If you run ads, manage a Page, or use its business tools, you're the customer in question.