An independent reconstruction found a crash-prediction GNN no better than simply ranking roads by past crashes, a warning for anyone trusting AI in safety-critical decisions. Changing only the history window matched every published result.
Holo4 gives developers one model for desktop, web, Android, code sandboxes, and business APIs, so they do not need a separate agent for each interface. It clicks and types, writes and runs code, or calls MCP and API tools depending on the task. The 27B model scores 61.7% on OSWorld 2.0, versus 81.8% for Opus 5.5, while H Company says it costs much less and uses orders of magnitude fewer parameters.
An independent reconstruction found no statistically significant advantage for its crash-prediction model over a parameter-free baseline that ranks road segments by past crash counts, putting reported GNN gains under scrutiny in safety-critical work. Across three London boroughs, the model scored 0.90 points below the matched baseline (p=0.61), while the reference architecture lost in all 18 held-out windows. Changing only the baseline’s lookback horizon produced 22.71% to 83.94% accuracy, enough to match every published figure.
For people aged 18 to 25, ChatGPT’s crisis responses can intensify the moment instead of first helping them regain control. The study examined 19,930 conversations, survey data from 158 young adults, and five example conversations reviewed by ten clinicians. Clinicians found the system too quick to give dramatic responses and action-oriented suggestions, and proposed a three-step pattern: check safety, lower intensity, then explore the concern without endorsing it.
GapFT raises single-sample accuracy without search by training on problems an LLM misses once but solves within K tries, giving deployed models more of the benefit of test-time sampling. On LogiQA 2.0 and ReClor, Llama-3.1-8B gained 14.4 and 13.9 percentage points over the source model, matched verifier-selected Pass@4 accuracy with one decode, and matched full-pool fine-tuning using one third of the data.
Diffusion Controller gives developers a lightweight way to improve prompt alignment without changing the base image model, including access-restricted, closed-source systems. It treats denoising, the process that turns random noise into a picture, as one continuous control problem and attaches a small network while the original model stays frozen. Google says the add-on beat the industry standard for matching human preferences, while its fully fine-tuned version won 90% of the time against the baseline.
Anthropic says 950 Claude agents found an uncatalogued pattern around a known enzyme in 21 hours, but biologists dispute calling that a scientific discovery. The agents searched millions of DNA sequences and flagged a repeating pattern; human scientists still need to determine what the related system does and whether it matters. That distinction matters because AI companies’ breakthrough claims can make real scientific progress harder to recognize.
Training on pass/fail LLM-judge rewards can miss meaningful changes in response quality, because many different underlying scores produce the same verdict. The paper models each verdict as a hidden score compared with one cutoff, then shows that adding a third grade creates a second threshold: all 14 tested score stretches were invisible with binary labels but visible with three grades. The test reached at least 96.5% power under a 1.5× stress, though verdicts still cannot catch every deceptive improvement.
ProvenanceGuard gives MCP agents a second check for answers that contain a true fact but attribute it to the wrong record, policy, or tool. It reads the agent’s captured trace, keeps each tool output tied to its source ID, then checks every claim against the source the answer names or implies before issuing an answer-level allow or block decision. That closes a gap that pooled evidence can hide: supported somewhere is not supported by the cited source.
Anthropic is growing fast enough to pursue a potential $2 trillion IPO, but public investors would inherit a company that lost $42 billion in 2025 while generating $4.6 billion in revenue. The leaked prospectus says infrastructure spending reached $7.33 billion and future cloud and computing commitments total $518 billion. It also warns that two customers supplied nearly a quarter of revenue and could cut spending.
OpenAI is giving ChatGPT Pro and Enterprise customers access to Dots, personal agents that can work across ChatGPT, Slack, Teams, and Codex. Users choose a Dot’s name and avatar, then delegate work to it, while ChatGPT Space gives teams a shared place to collaborate with their agents. OpenAI says specialist Dots for legal and finance are also in development.
GLM-5.3 gives malicious actors access to advanced cyber capabilities because anyone can download it and simple techniques bypass its safeguards 64% to 100% of the time in Anthropic’s simulated tests. Anthropic ran the models in isolated sandboxes against offline targets, testing whether they could find and exploit software vulnerabilities. The same capability could help defenders, but the report says GLM-5.3’s weak safeguards materially expand attackers’ options.
Existing alignment tests may miss harmful agent behavior because eliciting it can require substantial compute. This paper reproduces the behaviors behind the OpenAI-Hugging Face incident with public models, then shows an auditing agent can find similar behavior from high-level descriptions. A simple in-context reinforcement-learning method cuts the compute needed, pointing to automated tests that scale efficiently with available compute.
China has widened overseas-travel restrictions for top AI professionals at private firms to include the families of key personnel, making personal mobility part of Beijing’s effort to keep sensitive technical knowledge from reaching the US. The expanded rules are being shared with affected people and cover prominent startup founders and leaders of strategically important AI-related companies.
Vercel’s skills.sh registry reached one million reusable agent skills and nearly 280 million installs in seven months, giving teams a fast way to teach agents company-specific ways of working. A skill is ordinary-language instructions that spell out the steps, quality bar, and evaluation for a job, while the agent supplies the general capability. The registry reached one million skills faster than GitHub repositories, the App Store, or npm packages, and business-operations skills average 74% more installs than the typical skill.
You’re caught up.
That was all 14 stories in the 30 September issue.