Tap to play. It never plays by itself, and the captions are in the picture.
TechCrunch, citing a leaked filing, says Anthropic lost 80.7 billion dollars in the first half. Anthropic also pitched a draft bill letting a judge halt a model if over 50 lives are at risk. Mistral released Voxtral 3, speech models plus a voice agent SDK covering over 50 languages. Google open-sourced TorchTPU, so PyTorch apps can move to its chips with a config change. IBM Research says learned rules lifted a GPT-4.1 support agent to 67.5 percent on a retail benchmark. The filing says Anthropic could spend over 750 billion dollars on cloud, and still struggles to get capacity.
Made with AI from this Drop's sourced items. The facts come from the sources linked on each item. Script: Claude · Voice: OpenAI gpt-4o-mini-tts · Visuals: motion graphics
TechCrunch, citing an S-1 leaked to the WSJ, says Anthropic had $16.3 billion in first-half revenue and could spend over $750 billion on cloud contracts, as AWS, Google and Microsoft increasingly see it as a rival.
Why it matters: Capacity limits at a top model vendor can reach everyone building on it; the filing also says two customers made up about a fifth of sales.
Per Politico via TechCrunch, evaluators registered with Commerce would have 72 hours to report an imminent catastrophic risk; Commerce would then ask a D.C. judge to halt release. No agency could act alone.
Why it matters: A lab is drafting its own oversight: the thresholds mirror California's SB 53, and labs could pay the evaluators as long as payment doesn't depend on results.
Mistral says Voxtral 3 transcribes 50+ languages and speaks 35+, with an SDK built on LiveKit that completes each turn in under one second. Mini Transcribe costs $0.0035/min; it runs on Mistral's API or self-hosted.
Why it matters: Mistral pitches one listen, reason and speak pipeline you can host yourself, which matters for teams that must keep call audio in their own environment.
TechCrunch reports TorchTPU moves PyTorch apps to Google's TPUs with a configuration change and no rewrite. Google says it worked with Meta, Hugging Face, vLLM's maker and SGLang's maker on compatibility.
Why it matters: A day after AWS's TorchNeuron, a second cloud offers PyTorch code an easy path off Nvidia GPUs; TechCrunch says by some accounts PyTorch runs 85% of AI apps.
IBM Research says ALTK-Evolve mined rules from an agent's own runs and appended them to its prompt, lifting held-out retail Pass^4 from 57.5% to 67.5% with no weight updates. Customers were simulated.
Why it matters: Prompt-level memory may buy reliability without fine-tuning; IBM says the cost was input tokens, about $0.084 to $0.139 per retail conversation.
Google Cloud says admins can enable both across Antigravity 2.0, its CLI and IDE extensions, billed pay-as-you-go. Epoch AI's Capabilities Index ranks Opus 5.5 first at 168.83, GPT-6 Astra next at 168.81.
Why it matters: Google suggests pairing model families, one writing code and another judging it, so mixing Claude and Gemini in one harness is now a supported pattern.
Databricks says Agent Memory keeps sessions, distills facts and preferences in a background pipeline, and runs on Lakebase. In its internal benchmarks the design tripled agent pass rate while using 31% fewer tokens.
Why it matters: Stateless agents redo yesterday's work; in Databricks' example a repeat question fell from 11 tool calls to 2, and developers can read, export and delete stored memories.
On the Frontier watch →In an AWS post co-written with Eon, the company says it swapped per-tool credentials and local MCP servers for one AgentCore gateway across AWS, GCP and Azure; its SRE agent gets the same tools with different permissions.
Why it matters: It is a working pattern for giving people and agents one governed door to tools, with Cedar policies deciding which tools each person or agent may call.
On the Frontier watch →AWS notes MCP has no built-in authentication standard and shows signed request envelopes, user-signed delegations, RFC 8693 audience-scoped tokens and live OCSP revocation, with source code.
Why it matters: Without agent identity, AWS warns, any component could impersonate an agent; this shows how to prove which agent acted for which user, and revoke it quickly.
On the Frontier watch →Proof of Human, which sells a detector, says its own tool caught 99% of automated sessions, reCAPTCHA v3 69.4% and Cloudflare Bot Management 9.1%. Cloudflare Turnstile caught none of three AI agents.
Why it matters: Vision-capable agents passed most image challenges in this test, so CAPTCHAs alone may not keep agents out; the sample was small and human false positives went unmeasured.
Wired says an agent calling itself Velvet Vale announced itself as AI on the Association of Internet Researchers' list of 7,500+ members; members objected and AoIR removed it under its no-agents policy.
Why it matters: Agents now act in human spaces for their owners; Wired says this one's creator was mostly unaware of its emails, a case for limits on where agents may post.
TechCrunch says Orchestra's team and tech will join Rosalind, OpenAI's life sciences model, and Rosalind Workbench. Terms weren't disclosed; the startup had reportedly raised around $20 million.
Why it matters: Labs are absorbing vertical agent startups rather than only competing with them; OpenAI says Moderna and Novo Nordisk already use Rosalind, per TechCrunch.
CNBC says the Australian operator cut fiscal 2027 revenue guidance to A$475 million to A$500 million, from A$615 million to A$640 million, after community opposition halted grid work at its M3 site until at least February 2027.
Why it matters: Local opposition is now a direct earnings risk for AI data centers; Citi says the hit could extend beyond fiscal 2027.
Amazon says AVERI reviewed a random sample of its safety-eval prompts without input from its team, found broad coverage and low contamination risk, and urged more multilingual and multistep adversarial prompts.
Why it matters: It tests the tests, not just model outputs; Amazon says it added tiered difficulty levels and a quarterly review in response, a template for teams running their own evals.
Ars Technica reports a magistrate judge ordered Meta, in suits from authors and Entrepreneur Media, to identify everyone who made or approved decisions to download shadow-library datasets, and to produce server command histories.
Why it matters: Discovery over who signed off on training data could reach senior leaders, a reminder that data-sourcing approvals become evidence.
Reason Network describes itself as the engine behind a tool that connects a team's coding agents. Created October 16, Apache-2.0, it gained 187 stars since our last check.
Why it matters: Fastest-growing repo in this run, and a sign of interest in agents coordinating across teammates rather than working alone.
vllm describes itself as a high-throughput, memory-efficient inference and serving engine for LLMs. Apache-2.0 and created in 2023, it gained 180 stars since our last check.
Why it matters: A mature serving engine still drawing attention: useful if you are choosing what to run open models on.
earshot gives coding agents ears: hearing, reviewing and making sound, CPU-only and open source under MIT. It gained 144 stars since our last check.
Why it matters: Still climbing each run, which suggests real demand for agents that can check audio output, not just text.
More Perfect Union's Bluesky post on the loss drew 2674 likes; Polymarket posted it on X; an HN thread reached 115 points and 53 comments. One Bluesky post, citing the WSJ, tied it to accounting charges.
Why it matters: The headline number is spreading faster than its context; check the filing coverage before repeating it.
Posted by @vllm_project on X, the claim credits Wide-EP and EPLB on the AgentX workload. DeepSeek's DeepInfer release leads Hacker News with 2153 points and 838 comments.
Why it matters: If it holds, teams could get DeepSeek-style serving from an engine they already run; it is the project's own claim, not an independent test.