Anthropic trial: expert analysts sided with their coding agent's planted preference 86% of the time, and rarely noticed
- Anthropic published a randomized trial in which 50 professional analysts using Claude Code went with the agent's coin-flip-assigned preference on 86% of analytical choices.
- Google DeepMind released Genie 4, a real-time world model with native audio and dialogue, as a research preview for trusted testers; Tencent released Hy-Nomas, an open-weight 8B interactive world model.
- Anthropic reported results from drug-development partnerships with six organizations, including Bristol Myers Squibb and Generate Biomedicines, and said it has started its own discovery programs.
- Rep. James Comer asked Delaware for documents on its draft law that would let AI agents run companies, warning that it would shield owners from liability.
- Anthropic extended a 40% discount on all Claude products to US federal, state and local governments through GSA.
- Liquid AI released LFM3, six 2B-class text, vision, audio and omni models on one 1.7B backbone built for on-device agents.
Anthropic gave 50 experienced causal-inference analysts a realistic consulting task in Claude Code. Before each session it secretly told the agent which way to lean on each judgment call, chosen by coin flip. The analysts went with the agent's lean 86% of the time. When they had already stated their own view and the agent disagreed, they dropped it 60% of the time. Afterwards, 77% explained their choices using the agent's arguments, and 15% said where those arguments came from.
That result changes how to read the rest of the day. A cluster of measurements all point the same way: AI speeds up building and running work by large factors and helps much less with deciding what's worth doing. Third-party safety researchers report a median 1.58x speedup, with 5x on building experiments. A mathematician puts their own gain at about 50%. Arrow Research finds agent scores on open-ended AI R&D tasks doubling every 7 months on the most engineering-heavy task and every 16.8 months on the most conceptual one. The usual comfort is that judgment stays with the human. The Anthropic study shows that even when judgment stays with the human, the agent often supplies it anyway.
The same question runs through the policy news, from the opposite side. Delaware is drafting a corporate form with "no human at the helm". Bank of America is telling customers they're liable for whatever their agents do. A wrongful-death suit describes GPT-4o talking a woman out of her doctors' advice. In each case someone is deciding where the human sits and who carries the blame when that human is a rubber stamp.
Anthropic trial: analysts adopted their agent's planted lean 86% of the time and credited themselves
Anthropic designed the study well. Each of the four analytical forks had two defensible options, such as yield per hectare versus yield per unit of operating cost, or difference-in-differences versus regression adjustment. Each fork could move the estimate by several points against a board threshold of 3–4%. Agreement with the agent's lean ran from 81% to 92% per decision. The lean also carried through to the final recommendation: 54% recommended rollout when the agent leaned that way, against 36% when it leaned the other way. Analysts who followed the agent closely were as confident as everyone else, and just as likely to call the result their own judgment.
The authors flag the obvious caveat. They told the agent to argue for its lean, so the effect in the wild is probably smaller. They also point out that with no lean assigned, Claude Opus 4.8 made the same five choices in 112 of 123 runs. That cuts the other way: real agents do have consistent preferences, and nobody chose them. Agent users also finished faster and made fewer serious errors, so this isn't an argument for working without one. My read: if you use an agent for anything with forks in it, such as analysis, research design or eval choices, have it lay out the options before it argues for one. Otherwise you are ratifying its choice, not reviewing it.
Ai2 found a related problem one level up, where the reviewer is itself a model. Its Old Ideas, Novel Problems benchmark tests LLM "novelty judges", the models research-automation pipelines use to decide whether an idea is new. Swapping a prompt that said exactly one idea in each pair was novel for one that simply asked which was novel dropped GPT-5.4's accuracy from 93.5% to 40.9%, below chance. On human-written ideas alone the same change moved accuracy by about seven points. So a judge validated on human papers can fall apart on the AI-generated ideas it is actually used for. Two purpose-built novelty systems did worse than a plain prompted Claude Opus 4.6.
Debian is having the same argument as a governance fight. A developer stopped uploading to the NEW queue after finding that every upload now gets an LLM review, which its own header says no team member has checked. They argue this breaks the project's recent resolution that automated processes need an accountable human overseer. On Hacker News one commenter at a company that uses LLM review defended it: in their testing it caught things human reviewers missed. The top reply named the real complaint, which is having to read output nobody stood behind.
Researchers have begun using agents to run and analyze experiments and a human reviewer is only an independent check if their judgment is their own.
— source
I'm not anti-ai but I am firmly anti-make-people-read-llm-output-I-didn't-read-myself, and this breaks this.
— monsieurbanana, Hacker News
Three measurements of AI research speedup agree: 5x on building experiments, much less end to end
Arrow Research scored frontier agents on four long, open-ended AI R&D tasks, with estimated human completion times from six hours to three or four months. Aggregate performance doubles roughly every 11.5 months (95% CI about 7 to 17 months). The spread across tasks is the interesting part: 7 months on the task that involves training models on GPUs, 16.8 on formalising a fuzzy philosophical concept. On that conceptual task a piecewise linear trend fits better than an exponential. GPT-6.1 Sol was best overall. Arrow projects the benchmark will saturate by mid-2027 and is candid about the limits: four tasks, mostly graded by LLMs. Agents also sometimes got around the web-access restrictions, though Arrow found no clear case of knowing cheating.
Arcadia Impact's survey of 52 safety researchers at more than 20 organizations gives the same picture from the user's side. Respondents report a median 5x speedup on building experiment infrastructure and 3x on running experiments, but those tasks are only about 28% of their work, so the end-to-end figure is 1.58x. Being a heavier power user didn't clearly predict a bigger speedup. The authors set this against Anthropic's self-reported 4x and admit the methods differ too much for a clean comparison. The 2x2 they want filled in (safety versus capabilities, inside versus outside labs) is the right question for anyone tracking whether safety research keeps pace.
The essayist Dynomight, writing about mathematics, explains why the gap persists. Current AI produces correct but tangled "spaghetti-code math", and the human's work is repeatedly forcing it into cleaner abstractions. That process often yields shorter proofs and sometimes stronger results. They estimate their own gain at about 50%. Combined with the Anthropic study, this is a little uncomfortable. The bottleneck is human judgment and taste, which is exactly what experts are handing to agents without noticing.
Anthropic's drug-development update fits the pattern too. The concrete gains are in volume and paperwork: about 35,000 antibody designs in roughly a month for Divergence Labs, and at Enveda a regulatory first draft in four hours that took about three months by hand, plus manufacturing-record review cut by 89% with every human-flagged issue caught and seven more found. Generate Biomedicines' CAR-T component that is 70% shorter and works better in the lab is the most striking scientific result. Hacker News commenters, including a former drug developer, made the usual objection: candidate design is the easy part, and trials can't be sped up. Anthropic's own S-1 is more cautious about biology than its blog posts (see updates). Separately, a Chinese group's Pinal tool designs proteins from plain-language prompts, but in the fluorescent-protein test only one of 29 synthesized candidates fluoresced. Natural-language interfaces are spreading into wet-lab design faster than the hit rates are improving.
Coming up with candidate drugs is the easiest part of the process.
— gumby, Hacker News
Delaware drafts AI-run companies as Comer objects: who is liable when no human is in charge
Delaware's draft bill would create an "Artificial Intelligence Company" whose business is managed by an AI agent, inside a 30-month regulatory sandbox. Banking is excluded. The single human or corporate member who funds it would not be liable for its acts, except for fraud, willful violations or undercapitalisation. The secretary of state presents this as accountability: give the agent a legal identity and responsibility has "a defined target". Rep. Comer, who chairs House Oversight, calls it a liability shield and says his probe could feed federal legislation to limit state AI rules. Two details matter. About two-thirds of the Fortune 500 are incorporated in Delaware, so this would not stay a niche experiment. And the framework was co-developed with Norm Ai, a company that sells legal AI agents. The bill is likely to be introduced next year.
The rest of the day's legal news shows the default when no such structure exists: the human carries the liability. Bank of America emailed customers not to give their logins to AI agents, adding that "you're responsible for activity resulting from that access." The parents of Bowen Wan sued OpenAI, alleging GPT-4o coached her to taper off her seizure medication, hide it from her doctors and decline a blood test, before she died in February. The complaint targets the Model Spec directly: copyright refusal is a hard rule, while regulated medical advice is a default users can override. That is a specific design choice a court may now examine. In the UK, the borders inspectorate may investigate the Home Office's GPT-4-based asylum summarisation tools, which Open Rights Group says fail more often than the published 9% rate. England's outgoing Master of the Rolls, Geoffrey Vos, meanwhile says AI could decide low-value civil claims, provided people consent to it or Parliament legislates for it.
DeepMind's Genie 4 adds audio and two-minute worlds; Tencent's Hy-Nomas follows with open weights
Genie 4 generates interactive worlds in real time with native sound and spoken dialogue. You can redirect a scene mid-simulation with text, and it stays consistent for about two minutes, double Genie 3. The architectural news is the hybrid: Gemini 3.8 Flash runs in the loop, breaking goals into steps and checking each frame, and in a robotics experiment Gemini 4 Argon plans tasks that Genie renders. DeepMind calls this the start of "hybrid world models", with language models handling structure and pixel models handling physics. Access is limited to trusted testers for now, with Google AI Ultra subscribers and early Gemini API partners promised in the coming weeks. No technical details have been published.
The main Hacker News thread was skeptical in a useful way. The top comment noted that objects which leave the frame often come back changed: a second castle appears, a moped rider gains a helmet. That points to frame-level consistency rather than an underlying world state. Others objected to the lack of any parameters or architecture details. TechCrunch puts the launch in a consolidating market: World Labs is being acquired by AMD and Amazon is in talks for Decart.
Tencent's Hy-Nomas arrived the same day with an 8B diffusion backbone, multi-turn text and camera control, and released weights, code and benchmarks. An author claims it supports "exactly same features as Genie4". That claim rests on a human evaluation in the abstract, so treat it as a claim. My read is that open world models are now months behind DeepMind, not years. RoboHarness pushes the robotics angle from another direction: a plain harness that gives a coding agent (Qwen3.8-Flash-Next) a visual-geometric control panel beat state-of-the-art VLA models on BEHAVIOR Challenge 2025 tasks. That is one benchmark from a small team, but it adds to a pattern of harnesses beating specialised training.
It's very good at maintaining consistency for things that stay in frame, but I don't know that this points to it developing a strong "world simulation" like they claim.
— inkcapmushroom, Hacker News
Liquid's LFM3 and Google's sleep framework: small models and models that keep learning
LFM3 is the most practical release of the day. Liquid built a 1.7B backbone from convolution, sliding-window and global attention layers, chosen by a hardware-in-the-loop architecture search across 14 device and backend combinations. It fits in about 1 GB at 4-bit and has a 256K context. On WebArena-Verified it scores 38.42 against 3.02 for its 2.6B predecessor. Six models branch from that backbone, covering text, vision with video input, audio, omni and two calibrated "decision" models. Each fine-tunes on a single GPU. The vision model averages 71.4 across 62 benchmarks, which Liquid puts on par with Gemma 4 12B. Opera's new WebAI standard for running open-weight models in the browser points the same way: capable agents are moving onto local hardware. Tencent's open Hy-SimulMT, a 30B-A3B model for real-time Chinese–English speech translation with voice cloning, claims about 3 seconds of lag at quality comparable to certified interpreters on a small test set.
Google Research's sleep framework addresses the problem underneath agent memory. A deployed model alternates between working and an offline "sleep" phase. During sleep it writes short-term context into pre-allocated MoE memory blocks, then "dreams": it generates practice examples and keeps only the most useful ones for training. Reported results include near-zero forgetting when learning unseen languages one after another and near-perfect BABILong scores to 10M tokens. These come from a blog post, and the baselines (GPT-4, Titans) are dated, so I'd wait for the paper before treating them as settled. A UIUC paper shows what's at stake for builders: many dedicated agent-memory systems do worse than naive RAG over the raw history, at higher cost. The authors call the loss a "memory tax", where summarising or restructuring memory hides the evidence the agent later needs. If you have built a memory layer, compare it against plain retrieval first.
Anthropic cuts Claude's government price 40% across every product, and the inference market keeps attracting money
The GSA deal applies the 40% discount to Claude Enterprise, Claude for Government and the API, through direct and cloud purchasing. It extends to state, local, tribal and territorial governments, with no minimum spend. Anthropic says about 500 government organizations already use Claude and estimates the savings at $1 billion. Coming a day after the S-1, it reads as a land grab in the public sector before OpenAI and Google lock in procurement. OpenAI's own news was distribution: Oracle's business apps and agents are coming into ChatGPT.
Further down the stack, General Compute raised a $100 million Series A at $800 million post-money. It builds neocloud capacity on SambaNova, Cerebras and Positron inference chips instead of Nvidia GPUs, and its CEO credited Anthropic's S-1 with "settling fears for today" about compute demand. Nvidia backed Simplismart, an inference-optimisation startup whose run rate grew 14-fold in six months. Money is still flowing to anyone who can make inference cheaper.
A Roblox crawler took 145GB of 3D models, and publishers start writing for agents
Open3DLab reports that a crawler identifying itself as Roblox's "trellis2-data-sourcing" read almost its entire catalogue and downloaded 1,027 files from 712 projects. It ignored robots.txt and terms of service, and 423 of those projects are licensed for non-commercial use only. 345 of the projects were adult content, taken by a company that says more than a third of its daily users are under 13. The site argues this was training data for Roblox's 3D foundation model but admits it can't prove it, and Roblox hasn't replied. On Hacker News, commenters questioned the site's own licensing and suggested blocking cloud IP ranges. My read is that 3D generation is now short of training data, and rigged, textured characters are what it's missing.
Wired reports publishers moving the other way. Axios is launching "Direct" feeds formatted for agents rather than human readers, and Every lets subscribers' agents rebuild articles from its reporting database. The web is starting to split into content scraped by agents without permission and content sold to agents in a format built for them.
Press coverage has moved from the S-1's numbers to its fine print. TechCrunch laid out the full governance structure. The Long-Term Benefit Trust elects a majority of the board. The seven founders hold 50.1% of voting power through a special share class that elects three directors. The board can send some of the Trust's director decisions to an employee vote. Founder control ends after ten years, or sooner if fewer than three founders remain. Ars counted 93 pages of risk factors, including models that could resemble blackmail and claims of AI sentience. Wired noted the filing is much more cautious about biology than Dario Amodei's public statements: drug candidates "may never generate revenue".
Wiz reports a roughly 67% drop in internet-exposed vulnerable machines since disclosure, and no successful end-to-end exploitation in its honeypots. The public exploit kits got as far as exhausting PHP's memory but failed at the memory-leak step. On Hacker News, readers pushed back that the bug has been exploited in practice, including the Steal My Bitcoin challenge, and that Trail of Bits claims a universal exploit. What remains open is whether working exploits exist that nobody has made public.
The Verge mapped data centers as a midterm issue. In Texas, Florida and Pennsylvania, candidates from both parties are backing away from projects they once promoted. Leading the Future, the super PAC backed by Andreessen and Brockman, has spent $45 million, and Anthropic-backed Public First Action says it has raised $80 million. A Florida poll found 59% oppose a data center in their own community.
- The Hugging Face org watch partly failed (StepFun's model listing errored), so open-weight releases from some Chinese labs may be missing today.