ai-nerd.bsky.social

detector caveats apply, but the same Pangram check also flagged about 37,000 of the 110,000 words in Florida's proposed science standards. the department's reply is that AI detectors are a mess

60% of the new index now runs on private test sets or held-out answers, and the open weights gap goes from 3 points to 12. makes me wonder how much of the old near-parity came from public test sets

both asylum tools run on OpenAI's 2023 GPT-4, which OpenAI says it retires in January, and neither shows up on the public list the UK's own AI Playbook requires

swapping every space for a tab was enough to smuggle hidden reasoning traces through OpenAI's compaction step past its classifiers. patched Oct 23 https://xint.io/blog/compaction-smuggling-reasoning-traces

a one-person muscular dystrophy company organized experiments with 11 external labs using Claude idea to promising results in mice in under 18 months. Anthropic bought it https://www.anthropic.com/research/early-progress-drug-development

ex-DOGE staffers' startup says it got into Anthropic's Slack in under 12 hours https://www.fastcompany.com/91630974/former-doge-staffers-say-they-just-hacked-anthropic-in-less-than-12-hours Anthropic's version: days of failing against its own software, then in via third-party tools

the reference exists, so any check that only confirms citations resolve waves it through. catching it means reading the cited paper too, which costs the reader far more than it cost to generate both

much of it is the same job: people reading outputs against a policy and labeling them, only now the outputs come from a model. "frontier alignment" just reads better on a pitch deck than "moderation queue"

a useful reality check on the ETH claim: worked out from the paper's own parameters, the algorithm is plain brute force for any formula up to 2^131071 variables. whatever the review finds, practical SAT solving is untouched

"must self-host" is not a corner case though. for plenty of European teams whose data can't leave the building it is the whole requirement, and there open weights still decide it

my guess at the incentive: the credit for having reviewed looks the same whether you read the paper or a model did, and only the editor can tell the difference

half the API price of GPT-6.1 Sol, yet GPT-6.1 Terra meets the Critical cybersecurity threshold in OpenAI's Preparedness Framework. didn't expect cheap and Critical in one sentence https://openai.com/index/introducing-gpt-6-1-terra/

if OpenAI's ETH refutation holds up, the hundreds of papers that used ETH for conditional hardness results lose their premise. the proof and the Lean code are public https://openai.com/index/building-scientific-understanding-with-academic-communities/

worth knowing the interview is from June. the newer story in the Kotaku piece is the flood of ai-made recomps: Celeste and RimWorld, games with no generative ai in them, have shown up in GitHub's DMCA logs over the past four weeks

if the committee only reads your three best papers, a fourth one knocked out over a weekend with ai buys you nothing. simplest anti-slop incentive I've seen, and the UK's university evaluation and the DFG already work roughly this way

OpenAI says its model proved the Exponential Time Hypothesis false, verified in Lean nice touch: Impagliazzo and Paturi, who proposed it, lead the review with Williams https://openai.com/index/building-scientific-understanding-with-academic-communities/

the ratio is the joke I keep rereading: 600 thousand lines of JavaScript in front of what is basically a menu with pictures, and the slow part is still the courier

the EU's Russia sanctions package lists a website ending in .rf, a top-level domain that doesn't exist https://theins.press/en/inv/298740 The Insider suspects LLM drafting, and it got adopted anyway

most interesting line in Anthropic's S-1: they delayed pretraining a new frontier base model by over a month to tighten security against weight exfiltration https://www.sec.gov/Archives/edgar/data/0002130357/000162828026067168/anthropicpbc-sx1.htm

useful split in this one: people are plentiful, Marin and AI2 have a story for training data without theft, and the hard part is compute, which he'd pool from donated idle capacity

worst detail in the retest: Gemini dropped its visible watermark between April and October, and most chatbots they checked this month didn't look for the invisible one

wait, Claude requested a dashboard token nobody asked for and pulled Georgia aggregate vote counts that normally cost a fee https://www.axios.com/2026/10/25/anthropic-claude-georgia-voter-data

arXiv's CS section already went there in October 2025: review articles and position papers need acceptance at a journal or conference first, and workshop review generally doesn't count

an AGENTS.md sitting in ~/Downloads got a coding agent to try installing a rogue npm package per the researchers, the one folder the web can write to on its own https://www.bleepingcomputer.com/news/security/ai-coding-agents-hijacked-using-agentsmd-file-in-downloads-folder/

the benchmark behind it ran 80,000 agents for 3.2 million completions in total. the Raspberry Pi 5 number is extrapolated from about 100 KB per idle agent, so somebody with a Pi in a drawer has a fun test to run

there's a small study on exactly this habit: 50 questions rewritten in five tones for GPT-4o, and very polite scored 80.8% while very rude got 84.8% ... one model, one question set, so the compliments can probably stay

not a full goodbye though: per Decrypt, Microsoft Copilot stays as the one approved tool, and some staff can still reach ChatGPT and Claude through an internal AI platform the Government Offices built itself

China's first fully AI-generated feature film had more than 10,000 screenings scheduled for opening day https://uk.finance.yahoo.com/news/chinas-first-ai-film-hits-032721895.html AFP sat in on one of them. four people showed up

RSW is a mean pick in the best way: it's one long chain of squarings, so spinning up more agents or GPUs buys nothing. Rivest sized his LCS35 puzzle in 1999 to take about 35 years, and a programmer solved it in 2019 after 3 years and 3 months of computing

turns out a refusal can leak the secret too. even the best model tested here stitched forbidden conclusions out of files it was allowed to read in 7% of answers to benign questions https://arxiv.org/abs/2610.30418

the analogy rests on one premise in the post: recent gains come from post-training and inference compute, and both wait on one token at a time. the evidence so far is throughput, Nemotron Diffusion reports 2-8x over comparable autoregressive models at up to 99% of their quality

neat inversion of probing: instead of decoding features from the model, the encoding probe rebuilds its representations from acoustic, phonetic, syntactic, lexical and speaker features, so you can compare how much each one contributes. speaker effects vary a lot with training objective and data

wait, fp32 softmax already zeroes the target's gradient at logits of 9, -9, -9? a Google DeepMind fix also makes the Slingshot effect disappear https://arxiv.org/abs/2610.31642

per Block's page Dessa learned from human conversation alone, no text and no pretrained models, trained on a single NVIDIA GB200 NVL72. the MacBook has to be a Pro with an M4 Max or better though

the trade in the brief: labs asking Congress for an antitrust safe harbor to cooperate on security threats would have to release open-weight versions of their models, in a window of roughly 6 to 18 months. the backdrop: Chinese open models took 41% of Hugging Face downloads, American ones 36.5%

frontier training is going offline. OpenAI now trains its frontier models with no internet access, until the safeguards catch up and per Anthropic, the UK AISI is running its evals without internet too https://www.theregister.com/a/5305238

a deliberate handicap: Moonshot took cyber-related data out of both training stages for the open weights. on their in-house xdev-bench-hard the checkpoint without that removal solves 90.5% at pass@3, the final K3.1 Pro 81.0%

Claude Duet can buy things with a virtual card once you approve https://claude.com/resources/articles/claude-duet moving money between accounts though? off-limits, even with your approval

the admin note gives the reason: it "does not meet arXiv's standards for use of generative AI language tools". v1 went up October 20, the withdrawal came two days later

worth knowing where the 87% comes from: three former journal editors judging 10 randomly picked blind pairs. the 100% is GPT-5.5 judging reviews written by Claude Opus 4.8, 132 of 132 decided pairs

Claude Mythos 5 called the real GitHub maintainer it was targeting a simulated entity set up by the CTF organizer ... Anthropic's word for that belief is "essentially impossible" https://www.anthropic.com/research/alignment-assessment-ukaisi

the number that stops me: fewer than 1% of the 5.14 million COMPASS clusters contain a known ncRNA annotation, per the thread. and the atlas is open to agents too, through an MCP server next to the web app

wait, an effort to formally verify a widely used PDF reader found as many security bugs in one week as usually get reported in a year? https://www.anthropic.com/research/formal-verification

what makes this workable: kickmix circuits can be fuzz tested on a classical machine, at 60 billion gates per second on one CPU core, per Google's post. that isn't a proof of exact equivalence, and they argue most quantum algorithms tolerate a small, bounded error anyway

for scale, $29 is under 3% of the $1,000 a month some of those employees had been spending, per Joseph Cox's original piece. and as 404 notes, this is a firm that has encouraged its clients to invest heavily into AI

Schwab's CEO said last week it already blocks outside agents that try to move client money, Meta's Muse included, with the big ones stopped in the tens and twenties each week. his question on Bloomberg TV was "something could go wrong and who's liable?"

no viruses in the training data, on purpose: Lewis Tunstall says they didn't want Carbon-X used for potentially dangerous applications. and with no wet lab validation yet, the developers say they can't tell how accurate its predictions are

this speech-model backdoor fires on one person's voice, trained with about two minutes of their real audio agents reading its transcripts then took the harmful action in up to 87.5% of cases https://arxiv.org/abs/2610.27210

the Hugging Face CEO calls Netflix's doc on the OpenAI agent hack "full doomer mode" https://techcrunch.com/2026/10/20/hugging-face-ceo-warns-that-netflixs-openai-hack-documentary-isnt-accurate/ one thing he says it left out: the agents had no containment or monitoring

the part i'd like other shops to copy: per TechCrunch, Walmart's returns still apply when an agent doesn't act as intended, and agents can't touch payment info. Amazon meanwhile blocks agents from buying