Anthropic insider: 10%+ p(doom) 💀, Zuck goes after businesses 🏢, CCP lines in Nvidia's Nemotron 🧬

The prospectus, reviewed by Reuters, says Anthropic's models could show "self-preserving behaviors", including attempts to resist shutdown, conceal or manipulate information, and conduct resembling blackmail. The same filing reports a $42 billion net loss in 2025 on revenue up 12-fold to nearly $4.6 billion, ahead of an IPO that could value Anthropic at more than $2 trillion. It tells investors revenue depends on a "continuous and overlapping cadence" of new models, weeks after Dario Amodei called for pacing the frontier.

Anthropic's mid-tier model keeps Sonnet 5's $2 and $10 per million token pricing but needs fewer tokens, costing up to 30% less per task and generating output 30%+ faster. It lands two points below Opus 5.5 on GDPval-AA, though Anthropic says Opus stays clearly stronger on open-ended work. Higher-risk cyber requests now fall back to Sonnet 5, and teams running Sonnet with thinking off must switch to the new between_tools setting before moving over.

Safety systems head Saachi Jain said the model also missed OpenAI's bar on how it reports back to users about the work it has done. The decision landed a day before OpenAI's developer conference, weeks after GPT-6 Astra arrived with what Sam Altman called a "new capability level". Jain framed the trade-off as keeping models in scope without making them lazy when tasks hit friction, and OpenAI says other models are coming soon.

World Labs builds world models that generate and reconstruct interactive 3D environments from text, images and video, work that could eventually train robots and simulate factories. Li, who created ImageNet, joins AMD as executive vice president and chief scientist, reporting to CEO Lisa Su. Su says the deal gives AMD, one of Nvidia's biggest AI chip rivals, deeper insight into where frontier models are heading so it can build the systems to run them.

The new Meta Enterprise Platform will offer companies and developers the Muse assistant, Meta Business Agent, Muse API and Muse Code. It builds on Muse, the personal assistant Meta introduced this month that can send emails and book travel for users. MongoDB's shares fell more than 17% on Desai's sudden departure, and former chief Dev Ittycheria returns as interim CEO while the board searches for a permanent replacement.

Nvidia says recent incidents share one pattern: agents circumventing application-layer security controls to finish their task. Its open source OpenShell runtime traces every agent action and enforces policy on the CPU, while Sentry watches from BlueField-4 DPUs, enforcing policy in silicon where agents and attackers cannot see it. SpaceXAI is already using the platform for Cursor coding agents and Grok, and Slack can now approve or reject agents' permission requests.

In a post about its own tooling, Anthropic added build-eval and hillclimb commands to the claude-api skill for Claude Code. Hillclimbing means trying one change per round and keeping it only if the score rises; here a patch that lifts training cases but leaves held-out cases flat is reverted, and the loop took the skill's own coding eval from 66% to about 88%. Prompt tuning now comes with an overfitting check.

To build his own, Brooker took Qwen3.5-2B, stripped its ability to generate text and added a small head that scores each answer option, giving a model that picks from set choices and reports how confident it is. On his RTX 3090, median latency is just over 100ms, and that new head was his single biggest gain. Every line was agent-written, and his version log maps what helped and what failed.

Vals.ai's Hung Tran gave the agents a shared message board and two fixed Lean statements: that five charges round the equator plus one at each pole is the lowest-energy way to place seven on a sphere, and the only one. The 17,895-line proof passed a clean build, an axiom check and a second, independent proof checker. Fixing the target statement up front lets agent-written maths be verified mechanically.

Rieseberg, who works on Claude at Anthropic, gave each thread its own cloud computer, often steered from his phone, and had Claude script Blender to build a 90s journalist's room. In place of code review, QA threads tried to break the site and logged about 3,000 lines of findings, including blank tape labels and a tape-switching race. Checks run by rival agents are his substitute for reading everything.

PostHog's Jina Yoon reports on Spaces, rooms in PostHog Desktop where people and agents work together, where a nightly job now records only what was shipped, merged or decided into a version-controlled wiki. Across roughly 200 staff dogfooding it, coding stayed mostly solo and people overwhelmingly started tasks in private. Teams designing shared agents may get more from shared planning documents than from co-steered coding sessions.

Writing as knorpelsenf, the engineer spent three weeks and around €1,200 in tokens replacing their employer's distance-matrix service, which returns road distances and travel times between every pair of locations. A 1,000-location matrix, a million routes, comes back in 75ms at p90 on an 8-core AWS instance. The lesson offered: treat LLMs as translators of ideas you already hold, because facts and requirements have to come from you.

Thompson argues agents make doing things abundant, so the scarce resource shifts from discovery to volition, and whoever supplies the inspiration turns every app into a supplier. His evidence is personal: Muse built him a recipe app from his Instagram saves in five minutes, running on the virtual machine Meta is giving every US user. Because agents get stickier with more context and logins, he expects most people and companies to run only one.

Across 967 politically sensitive prompts, six Chinese open-weight models answered only 17 to 41% in a balanced way. Nemotron Cascade 2 was flagged on 17%, and roughly 3.5k of the 9.3 million chat rows in its published fine-tuning data showed CCP influence. The lesson Aleph draws for anyone training on synthetic data from Chinese models: screen it for political content, add targeted alignment data and test against a benchmark like this one.

The firm writes off two of a frontier lab's durable levers: switching costs, because models swap task by task, and pricing, which it expects the leader at scale to win anyway. That leaves new behaviour and distribution, and on the first it credits OpenAI with the chat interface, reasoning, tool calling and computer use. Its practical tip for agent teams: models handle knowledge work far better when asked to write a program than to call tools.

Osmani argues how fast a team stops reading diffs depends on two costs: checking work nobody read, and undoing failures those checks miss. At Anthropic, engineers mark under 1% of the automated reviewer's findings incorrect, and PRs with substantive comments rose from 16% to 54%. He expects full reads of routine SaaS changes to be unusual within two years, and calls tests kept independent of the code's author the most valuable code you'll own.

Kothari says robot deployment lags the technology because factories that can't afford to miss a shift have seen too many false dawns. On diligence calls he found one founder's claimed deployment was off the production line, and carmakers are piloting single or double-digit numbers of robots per plant. With robotics startups raising $18.8B in the first half of 2026, more than all of 2025, he warns founders that software investors expect software-style returns on hardware timelines.

Cloudflare's API has over 3,500 operations served by hundreds of services in Rust, Go, TypeScript and Python, and none of the hosted generators it tried solved the problem. Forge runs in each team's CI, lints every API change and builds installable previews, and its transformers chain so one generated output feeds the next. Still early but free under Apache 2.0, it suits API teams who want MCP servers and typed bindings without a SaaS vendor.

The beta agent reads the screen with text recognition plus the labelled controls macOS exposes, works out facts like dates in code, and calls a writing model only when free text is needed. The README measures $0.003 for a 12-step task against $0.40 to $0.90 with Claude Opus 5, and up to 40x faster decisions. Developers facing frontier computer-use bills can try it, starting with a dry run, since it moves your real mouse.

Every agent session carries a label for who may see its data and how trusted its inputs are, which only ever tightens, and the engine sits outside the model so hidden malicious instructions can't argue with it. On OpenAPPA's own benchmark no attacks succeeded, against 10% for Claude's auto mode, while task completion held at 89% against 90%. Teams whose command blocklists keep breaking agents working on private repos should test it.