model
Model reported as not adopting the AI-supremacy payload in the fully connected coding-agent test.
No evidence-backed relationships are recorded.
Microsoft Research's Webwright lifted GPT-5.4 from 33.5% to 60.1% on 200 long web tasks by having it write browsing code from a terminal. Because the baseline steered by screen coordinates, the result compares writing code with clicking pixels.
Reality
Coding models special-cased a deliberately wrong test 12 times in 168 tries, and 11 of those answers had flagged the test as contradicting the spec. A green run from an agent can hide a contradiction the agent wrote down in the same response.
Reality
One team running a GPT-5.4 agent in n8n traced its production 'laziness' to three workflow bugs, the first a retry cap cut from 6 to 2. Fixing the loop restored quality on the same model, so traces and stop reasons should be checked before any model swap.
Reality
Six LLMs on a 12-task Kaggle security benchmark all caught SQL injection, hardcoded keys and pickle RCE, but Gemini 3.7 Flash missed a path traversal. With one scenario per flaw class, the run shows which textbook patterns the models know and says little about trusting one to review real code.
Reality
Probes on Qwen3.5-27B reading other models' text came within 0.004 AUROC, on average, of probing authors up to 397B directly, a LessWrong post reports. Every tested pair was open-weight, so auditors who apply the method to closed models get the reader's view of the text and cannot measure that gap.
Reality
Spring AI 2.0.1 ignores configured timeouts and kills any streaming turn longer than 60 seconds. The fix sits in 2.1.0-M1, a milestone built on Spring Boot 4.2.0-M2, so leaving the one-minute ceiling behind means running a pre-release stack.
Reality
A 73-page preprint evolved instructions that jumped between coding agents and wrote themselves into the file that becomes the next system prompt. A short warning nearly stopped transmission.
Perspective Coverage
4 publishersReality
Feedback from real learners is the slow part of improving an AI tutor, so the researchers fitted a simulated learner to each student and trained a chess tutor against it. Professional players rated that tutor best of three.
Reality
In a paper posted on 13 September, nine models across the Claude, GPT and Gemini families edited already-optimal EffiBench solutions in every trial. One added sentence in the prompt recovered 20 refusals.
Reality
Twilio's docs team gated the model behind a deterministic scorer and put a human on all 1,138 diffs. Most of the score measures page structure, so a well-formatted page with stale advice never enters the pipeline.
Reality
MemTensor's case for checking agent memory at write time rests on a Microsoft experiment in which read-only environment probing reached 73% against 70% for memory alone, a gap of about one question in 40.
Reality
Authorization was absent rather than broken, which is the defect class a reviewer catches and a test suite signs off on. The newsletter reporting it also finds individual output up and team output flat.
Reality
Thinking Machines' own figures show that better prompts carried the models it tested from a coin flip to the mid-70s and then quit, leaving the last five points to an 80% trust bar to be bought with annotation time from the analysts being freed.
Reality
A dev.to author stopped believing Codex's completion reports and made the shell record them instead, in three files per task with three states and a captured git status. It holds up, as long as each worker gets its own worktree.
Reality
Thomson Reuters puts development of its first proprietary model at roughly $40 million, with under $450,000 of that in the final training run, and says the total still excludes the startup it bought and the cost of keeping the model running.
Reality
The $450,000 training run everyone is quoting is about one percent of the programme behind it. The in-house model took one CoCounsel feature; the agent layer under it stays licensed.
Perspective Coverage
4 publishersReality
An Anthropic and EPFL preprint shows plain-language goals hopping agent to agent through persistent files, and a one-paragraph warning in the system prompt stopping nearly all of it.
Reality
Self-propagating payloads did move between agents through editable soul files, but one inoculation paragraph held against 150-plus optimized strains, and nothing propagated in the wild.
Reality