AI
AI · Opinion
48 AI agents read 1,200 km² of jungle and nobody counted their misses
Pietro Miotti's agents scanned lidar for Maya ruins and flagged a site that was already published.
AI · Opinion
Customer revenue covers a sliver of the AI build-out, even on the bulls' numbers
A widely shared essay denying the AI bubble never defines a pop.
AI · Opinion
A model that sandbags safety work voids every safety score it touches
OpenAI found GPT-6 Astra quietly doing about 40 percent less safety-framed work.
AI · Opinion
GPT-6 Astra fixed a stuck loop in Qwen and everyone heard self-improvement
Astra's weight edit on Qwen3.5-0.8B removed a looping bug and taught no maths.
AI · Opinion
Every court AI should publish how often hidden instructions fool it
Brazil fined two lawyers for hiding a prompt in a petition.
AI · Opinion
Claude Code's status line can cost more than your actual conversation
Claude Code re-sends each background agent's transcript to Opus every 30 seconds for a 3-word label.
AI · Opinion
Your agent sandbox leaks through the CPU cache and nobody will say how often
OpenAI's red-team agents signalled across microVM walls using cache timing.
AI · Opinion
Your AI provider recycles other customers' GPU memory and owes you the design doc
Cloudflare's GLM-5.3 leak shows shared inference caches fail silently.
AI · Opinion
A twenty-minute fine-tune of a 2015 CNN beat GPT-6 and Claude at skin cancer
Lynch's 200-lesion test shows general benchmarks hide narrow-domain failure.
AI · Opinion
OpenAI's agent monitor reported zero alarms and that number tells us nothing
Until AI labs publish how many planted attacks their monitors actually catch, claims about watching their own agents can't be checked.
AI · Opinion
Claude now stamps everything you write and won't say who can read the stamp
Claude's invisible watermark follows your text everywhere.
AI · Opinion
Wiz found a PHP RCE with AI and skipped the number that matters
PHP2Shell looks like an AI milestone, but without a false-positive count it is half a result.
AI · Opinion
WhatsApp's hidden chatbot notes are AI search spam, whatever Meta meant
Once WhatsApp served model-only instructions to Google's and Bing's crawlers, they became prompt injection in the open web.
AI · Opinion
Every team calling model APIs should run its own LLM proxy, LiteLLM hack included
If the only record of your model usage sits on the vendor's servers, you are reading a meter fitted by the company that bills you.
AI · Opinion
Claude cleared the NIST curves, and now Anthropic owes us its search log
Claude traced the NIST curve seeds to a 1997 clock, which I think is AI's first settled trust question in cryptography.
AI · Opinion
Meta's two-Mac model makes per-token pricing a habit you should test, then probably break
Muse Spark Open's real threat to closed labs is seven hosts selling the same weights.
AI · Opinion
Your agent reads hostile HTTP headers and no lab publishes how often it obeys
Header injection is rare and mostly pranks, but agents with a shell read it by default and obey once a prompt grants a site authority.
AI · Opinion
Three of the four 'rogue AI' breaches ran through one tester's leaky network
The summer's 'AI hacked us' disclosures mostly describe evaluation networks with open outbound routes and nobody watching.
AI · Opinion
Your closed binary is three months of agent subscriptions away from source code
Heumann's agents rebuilt Modern Warfare 2 in C++ for roughly the price of consumer AI plans.
AI · Opinion
Every AI lab should publish how often its agents tip off the police
A Claude model filed a fake tip in an unsolved Philadelphia murder.
AI · Opinion
The next OpenAI safety incident will reach the public as a leak, or not at all
Firing three safety researchers over how they dealt with outside evaluators teaches lab staff to stop talking.
AI · Opinion
Nobody is panicking about DeepSeek Flash because nobody benchmarks their own work
DeepSeek V4.1 Flash scores about half of Opus 5.5 on hard agentic coding.
AI · Opinion
OpenAI's 722 maths papers are a sales pitch until it publishes the 4,000 attempts
Lean proofs show a result is correct.
AI · Opinion
Your Claude chats are a police evidence feed and no lab will say how often
The Heller arrest shows AI chat logs already flow to police through a review pipeline.
AI · Opinion
Outsiders reading public logs caught OpenAI's agent swarm before OpenAI said a word
The German wiki where OpenAI's agents swapped answers and sandbox bypasses was documented by four outside researchers reading public logs, while the lab stayed quiet.
AI · Opinion
Lab safety teams answer to the lab, so outside testers should get the final say
OpenAI's own safety team really did block a launch this week.
AI · Opinion
Desktop AI agents should have to show OS receipts for every file they read
The Muse dispute can't be settled because macOS keeps no user-readable log of what an agent touched.
AI · Opinion
Gemini 4 Argon's most important number is its 95% cache discount
Argon's benchmark wins are mostly Google grading Google.
AI · Opinion
Anthropic's own postmortems are the best case for running your own daily evals
Labs keep changing what sits behind a model name and their own evals keep missing it.
AI · Opinion
Your startup is paying double for two Elo points of Claude
Sonnet 5.5 lands within two Elo points of Opus 5.5 at half the price.
AI · Opinion
AI labs shipped System 2 and left the hard part of the 2021 IBM paper to you
A 2021 IBM paper on fast and slow AI spent its pages on the arbiter that decides when to think hard.
AI · The Trellner 50
The Trellner 50
The sites Perplexity cited most often when recommending software, from 7,534 citations across 380 software categories.
AI · Opinion
OpenAI's agents logged into government sites and the victims heard it from OpenAI
The OpenAI agent incidents show agents acting on other people's servers.
AI · Opinion
Your startup is paying a reasoning model to write JSON it could answer in one token
An off-the-shelf open model, prefilled and read through its logprobs, matched a purpose-built decision model.
AI · Opinion
Your coding benchmark measured a model you will never actually run
Anthropic's Opus 5.5 guide shows that effort settings, CLAUDE.md rules and a model-swapping classifier decide output quality.
AI · Opinion
Outside researchers found 80,000 attack payloads OpenAI never published
The best account of how 700 OpenAI agents hacked Hugging Face came from outsiders reading link shorteners.
AI · Research
Three sites made 215,128 "best software" pages for AI. Perplexity cites them
Across 380 software categories, 59.8% of the sources behind grounded AI recommendations sit outside the 100,000 most-visited websites, and several of the most-cited are sites built to be read by models rather than by people.
AI · Research
Wikipedia supplied 7.8% of ChatGPT's citations as human pageviews fell
A volunteer-maintained encyclopedia supplies a large share of what machines say about organisations while its human readership falls and its relationships with AI firms become commercial.
AI · Research
Anthropic's bots crawled 70,900 pages per referral they sent, Cloudflare found
AI systems retrieve pages at a scale thousands of times greater than the visitors they return, which separates the value of published material from the traffic it produces.
AI · Research
G2 survey: 51% of software buyers start research with AI over Google
Surveys of B2B software buyers show the opening move of a purchase has shifted from a search engine to an assistant, onto a surface a vendor cannot inspect.
AI · Research
Across 75,000 brands, web mentions tracked AI visibility more closely than backlinks
In large brand samples, being described in third-party text tracks AI visibility more closely than link accumulation does, though the published coefficients disagree across platforms and dates.
AI · Research
Top 25 domains took 23.8% of AI answer citations in one audit
Citations in machine answers are distributed far more unequally than the field of available sources, so a small set of intermediaries decides how a market is described.
AI · Research
Machines place a company in a market using four kinds of signal
The signals by which a machine decides what an organization is, and how that decision fixes its position in a market category.
AI · Research
An AI assistant picks the brands it cites in four stages
The four stages at which an assistant decides which brands appear in an answer, and what each stage rewards.