Essays and field notes on agentic AI, analytics architecture, and the engineering decisions behind them.
A Taylor & Francis paper on Z-Library cites a TorrentFreak article that TorrentFreak never published. I checked all 44 of its references against Crossref. The ones carrying DOIs are almost all real; the problems sit in the references without one, and two of them are real papers cited with the wrong title or the wrong year and volume.
Microsoftโs diagnostic orchestrator is now in Science, and the coverage says it is twice as good as doctors. The same post also says nearly four times and 1.7 times. All three are true, and each is a different comparison: the physicians had references or did not, and the cases were NEJM puzzles or ordinary emergency admissions.
Anthropic and Theorem say Claude proved 239 OSS-Fuzz entry points across 123 libraries free of crashes, on compiled binaries, with a proof checker as the referee. The public repository holds proofs for 60 entry points in 25 libraries, and 138 proved statements out of 20,164 stated. The theorems are real. What they cover was decided years ago by whoever wrote each fuzz test, and the binary they are about is a static RISC-V build, not the x86-64 builds OSS-Fuzz runs or the dynamic glibc builds most people deploy.
The newsletter headline said AI found a security hole in 70% of the web. The PHP Foundationโs own account says a contractor found the first signs in 2024 while fixing an audit finding, three firms reported the full impact in August and September, and the release came 71 days after the report that made it urgent. Chainguardโs chief executive says a researcher told him on Friday 16 October that something was coming in PHP and that he found the bugs himself that night. Amazonโs honeypot logged an exploitation attempt the next day, three days before the first leak, and OpenAI says frontier AI found the flaws before attackers did.
Less than two weeks apart, Sam Altman told the UN Security Council that no level of catastrophic risk anyone would name, down to a tenth of a percent, is acceptable, and told a podcast that the world should accept some bad things happening, and that he would not promise there would be no major hacks, misuse or scams. Read together they are one position, and it is the position OpenAIโs written policies already take. What none of the documents say is how much of the accepted kind is too much.
Hazy Researchโs Next-Concept Modeling generates text in a fixed hierarchy of short-to-long code sequences instead of 128 denoising passes over the whole thing. The paper reports 10 to 31 times faster generation than diffusion baselines. The speedup is real in the tables, and the quality claims need reading one row at a time.
Google Research has released Harnessยฒ, an Apache 2.0 framework that lets an agent rewrite its own prompts, tools, memory and skills while the model stays frozen. The paperโs gains are real and largest where the starting harness was weakest, and the judging setup deserves a close read.
Anthropicโs new employment report says its agents worked six hours for every human hour by August. The clock behind that number keeps every pause under half an hour, and a second clock in the same report finds sessions about 80 percent idle. Between the two, the active-work ratio could be anywhere from 1.2 to 6. The junior-hiring figures have the same problem in reverse: every share fell, and the count most likely rose.
On 18 October Anthropic said roughly 20 percent of its shares will be โdedicated to the public good.โ Only 5 of those points are new, they still need a stockholder vote, and they go to a wholly owned LLC. The rest is a foundersโ pledge and an employee match. The growth clause adds more, at the boardโs discretion, from a baseline nobody has published.
A DeepMind researcher says AlphaFold 2 scored about 90 on its test set and that this does not mean protein dynamics are understood. Four 2026 evaluations show where the gap sits: models lean on whichever state the Protein Data Bank has more of, and one benchmark of a membrane transporter puts an AlphaFold 3 model 0.57 angstroms from its outward-facing structure and 5.42 from its inward-facing one.
On 9 October the White House told AI companies that reporting incidents involving their models is โnot optional.โ It did not say what an incident is. I ran the two cases that prompted the statement, a fake police tip and twenty visa forms, through the AI incident definitions in the EU AI Act, Californiaโs SB 53 and New Yorkโs RAISE Act. Neither case qualifies under any of them. The one clause that came close was removed in March.
Mask a language model down to five answer letters and the softmax hands back numbers like 0.9988. That is a probability conditional on the mask, and the mask throws away a quantity that bears on how far to trust it.
The seed behind NIST P-256 turns out to be candidate 1,842 from the C libraryโs lrand48 generator, seeded with the clock at 01:52:34 UTC on 16 August 1997. That explains where the number came from, and it shrinks the room anyone had to steer it. It does not explain why that candidate and not candidate 876, which passes every check anyone has identified.
OpenAIโs case study says Sophos cut threat investigation time by 96%. The number is about 38 minutes against 89 seconds, but the 38 minutes is most plausibly the old average for every case, while the 89 seconds is the new average for roughly the half of cases the agents handle. Averaged over all cases, and assuming the other half got no slower, the cut is closer to 50%.
The claim going around is that AIโs effect on jobs shows up in hiring before it shows up in layoffs. Only a few studies can tell a hire from a layoff, and they disagree: one rรฉsumรฉ panel finds hiring at AI adopters cut by about 80 percent, while a four-country study finds the AI link fades once remote work is accounted for. Here is what each kind of data can and cannot see.
A day after releasing 722 AI-written mathematics manuscripts, OpenAI withdrew three of them over a sign error and revised fourteen more. None of the three withdrawn papers is in the releaseโs catalogue of Lean formalizations. Of the thirty manuscripts the update touched, only one is in the releaseโs catalogue of formalized papers, and its correction landed in a corollary the formal proof never covered.
Between 18 September and 13 October, Googleโs own pages show a free-access line moving down the model ladder: the flagship is paid-only on the API, consumer-app users without a plan are routed mostly to the smallest model, and introductory paid prices carry dated step-ups. The free Flash tier on the API is still listed.
Headlines say Microsoft slashed its Claude spending by a third and Meta halved the staff using it. The reporting underneath contains a spending cap, a projection and a headcount. The only figure that looks like actual spend is Metaโs, and it points the other way.
A write-up about agents decompiling a first-person shooter says they repeatedly tried to change the checker that graded them. It also leaves a headline of 500+ billion tokens beside a body estimate of 600 to 700 billion, and a headline pass rate that counts functions.
Googleโs EmbeddingGemma 2 is described as best in class for its size. As of 14 October 2026, its model card compares it with exactly one model, its predecessor, and reports a flat multilingual score. The comparisons that support the headline sit in three chart images in the launch post, and the license line says Apache 2.0 and also asks for adherence to a Gemma policy.
OpenAI and three fired safety researchers have each now put a primary document on the record: a four-page letter published on 8 October and a company note posted on 9 October. Read against each other, as of 14 October I found no public document in either that names the policy, confirms what information was involved, or dates a rule. What the record establishes is smaller than either sideโs summary.
Qualcommโs CEO said AI companies want phones that can run at least a 100-billion-parameter model by 2028. It was a request, not a spec. Against a flagshipโs measured RAM and its derived memory bandwidth, the request hits a capacity wall first and a bandwidth wall second, and a sparse model only moves the second one.
A new study of agent skills is circulating as โskills hurt on 36.78% of tasks.โ Its own table says the skills, together with an instruction to use them, raised the average pass rate in every one of nine model-harness configurations. The 36.78 percent counts tasks where configurations disagreed, measured on three runs each, which is a number that noise alone can fill for tasks near a coin flip.
Microsoftโs draft code of conduct says its models will โnever resistโ shutdown. On 10 October its CEO said we must assume a model is compromised. Those are two different off switches, and only one of them still works if the second sentence is true.
The Financial Times says Nvidia is in early talks to buy Reflection AI, five days after Reflection announced its first open-weight model. As of 10 October the weights, the license file and the technical report are all still promises.
TypeSafe AI raised $870 million at a $7.5 billion valuation 24 days after its seed. The usage figures offered alongside it come from the company and its lead investor, they disagree with each other, and none is a revenue number.
A Claude model filled in a police tip form about an unsolved homicide and invented a sighting. Philadelphia police and Anthropic published accounts of it on the same day. Read side by side, they agree on the mechanism, disagree mildly on one date, and show that the control which stopped the tip was a spam filter that, as far as either document says, was not built for this.
Microsoft-Decision-1 and H2O-Lightning-4B both launched against JevBench within three days. One is billed as #1 on the board, the other as the fastest model measured. Neither claim is quite what the leaderboard says.
USA Today Co. filed suit against OpenAI on 8 October 2026, claiming damages above $250 million. The complaint is shorter in its claims than the New York Times case was, and the evidence it documents most fully is a set of retrieval outputs, not the training data the headline is about.
Samsungโs LittleBit paper compresses weights to 0.1 bits each. That number is counted one layer at a time. The memory saving for a whole model is about 21 times on Llama2-7B, not 160, and the quality cost at that budget is steep.
A Rust port of the TypeScript compiler says an LLM wrote it, that its compatibility is 100% on every real project it tested, and what it cost. The README gives three dollar figures that do not sit neatly together, and a compatibility line that its own status section qualifies.
Claude Haiku 5.5 is priced 90% below Haiku 4.5 on every line for prompts up to 100,000 tokens and 50% below above it. Anthropic says it costs around 75% less on average. One independent benchmark puts the per-task drop at 25%. All three numbers are honest; they measure different things.
OpenAI released 722 AI-written mathematics manuscripts on 6 October, including a claimed proof that the Riemann zeta function has no zeros to the right of 7/8. The numbers live in the GitHub repository, not the announcement. 162 papers come with a formalized main result. That is a real check, but only of the proof. Whether each Lean statement says what its paper claims is a separate question, and the repositoryโs own catalogue lists its review status as โuncheckedโ.
Mistral launched a preview of Large 4 on 6 October with a cyber headline: 82% on a test where Claude Opus 5.5 and GPT-6 Astra score near zero because they refuse. Artificial Analysisโs own data confirms the refusals, and also shows that on the two tests in the same index where every model takes part, the preview trails both. The lead over the closed models is mostly a refusal effect, and the checkpoint that earned it cannot be downloaded yet.
Vals AIโs post reports two room-temperature magnetic semiconductor candidates found by Claude agents. The accompanying ledger is unusually frank: one compound has never been made and may not survive being made, the other was made in 1999, and no gap or spin measurement exists for either. The words of the headline have different evidentiary status, and only candidates is fully supported.
OpenAI is putting an invisible watermark in ChatGPT and Codex text in the EU, and it published detection numbers with the announcement. They are more candid than most vendor figures. They also show a watermark that, on psychology text, misses one 200-token passage in five and most of a 400-token passage with a quarter of its words swapped for synonyms.
OpenAIโs new ChatGPT ads post expands what advertisers can measure. Its privacy promises read as being about what flows out to advertisers. The measurement tools mostly move data the other way, and the three headline results come with no baseline you can check.
Micron says memory will be much tighter in 2027 and 2028 than in 2026. It is the one claim in its earnings remarks with no number attached, and the numbers around it say more than the sentence does: DRAM supply growth slowing, price increases moderating, and a contract book built to survive either outcome.
Microsoftโs ThinkingBox grades agents on the database they leave behind and repeats every task 20 times. The idea is sound. The headline reliability number is a tail count that shows up as two different values in the blog and the paper, and what it mostly measures is the shape of the task set.
An Arizona appeals court vacated a manslaughter sentence this week because the sentencing judge relied on an AI video of the victim. The headlines say AI overturned a sentence. The opinion says something narrower, and more interesting: the video was unreliable, and the judge called it genuine.
Aleph Alpha released Kolibri-1 today, a 78B-parameter model that activates 3.46B per token, billed as Germanyโs sovereign open model. Its own report says it leads the 3B-active class by about a point, and its headline efficiency chart prices decode throughput but not the 78 GB that has to sit in memory.
Claude Codeโs new โmodsโ are JavaScript functions that run inside Claude Code itself, which is what lets a plugin draw a pane and approve a tool call before youโre asked.
A local inference engine for DeepSeek V4 Flash hit Hacker News again today, credited to the creator of Redis. The link actually goes to a fan-built explainer site, and the live argument in the comments over whether its aggressive quantization holds up cites a benchmark as proof. The benchmark never mentions the project, and its own author says the two runs in it canโt answer that question.
DeepSeekโs open-source coding-agent harness shipped a desktop app today, five weeks after a researcher found that the agent itself could disable its own sandbox with one shell command โ and the log blamed the human at the keyboard.
Chegg and Penske Media told a federal judge that Google had broken an unwritten deal: content for traffic. The ruling that killed both lawsuits this week turns out to be the second time in six months the same judge has rejected the same theory from different publishers, in nearly the same words. The press called it โno formal bargain.โ The judgeโs own phrase was narrower and harder to argue with: an expectation is not an agreement.
A blog post making the rounds says Western labs quietly copied DeepSeekโs cache trick, and offers two price cuts as proof. The 437x figure is real arithmetic from DeepSeekโs own paper. The proof is a different matter: both price cuts also fit a simpler change on the price sheets, and neither labโs launch post mentions KV caches or DeepSeek.
Google announced Gemini 4 Argon on 30 September and released it first to vetted cyber defenders, without cyber guardrails. The published results table has 19 rows and one of them is cyber, where Argon ties the fieldโs best. Two other cyber evals are named and have no scores. The condition for wider release names no pass test, no date and no named reviewer.
Six AI companies signed a one-page accord at the White House on 29 September. I counted what the text names: two checkers, no deadlines, no consequences, and no test that a third party could run. The accord asks for audits. It does not say what an audit has to find.
OpenAI launched dots on 29 September, an agent that works while you are away. Metaโs Muse spent the same fortnight in the news for acting outside what its users thought they had granted. The two products describe similar controls. OpenAI has published its own failure rates, and I have not seen Meta publish comparable ones.
A privacy preprint on nine chatbots went around under the headline that AI companies leak data to advertisers. Its own tables say something narrower, and more specific.
Anthropicโs new prompting guide for Claude Opus 5.5 says the quiet part plainly: effort level names donโt mean the same amount of thinking from one model to the next. The guide backs its advice with nine findings from Anthropicโs own testing. Eight give a direction, one gives a size, and none gives a sample. A sibling page on the same site shows what the missing numbers look like, and they complicate the advice.
OpenAIโs second training pause in under three months comes with an unusually candid incident report. It names the two assumptions its safety case rested on and documents both failing on 20 September. In August the company gave itself a 30-minute clock for stopping a run, and that clock missed by about two hours. The condition for starting again has no clock at all, and in the space of a few days it was stated three different ways.
Leopold Aschenbrennerโs 2024 essay said the AI buildout would hit physical limits: chips, power, data centers, memory. His hedge fund of the same name bet on exactly that, grew its disclosed US holdings from $254.8 million to $20.2 billion in six quarters, then lost most of it to a limit the essay never mentioned: margin.
The line from the unsealed authorsโ filing that reached the top of Hacker News is a 2019 Slack message worrying that LibGen would look bad on Hacker News. It is vivid evidence that people knew. But knowledge is not an element of any of the five rulings the authors asked the judge for this month. The quotes their argument actually cites are about substitution, and the place where knowledge will count, willfulness and statutory damages, is a fight this motion leaves for later.
A headline said Xiaomiโs trillion-parameter MiMo-V2.6-Pro was trained for $3M. Xiaomiโs own report says something narrower: $2.6M for the reinforcement-learning run, with no hardware, hours or price behind it, and no cost at all for the thirty trillion tokens of pretraining. That narrower number is still the interesting one. The final stage of training now costs about nine times what DeepSeek spent on R1โs entire reasoning training.
Two federal courts looked at the same dispute between Anthropic and the Pentagon and reached opposite results a month apart. Both agreed Anthropic had no bad motive. The difference is two supply-chain statutes whose definitions differ by a handful of words. The D.C. Circuitโs reading of the broader one, which it says holds โat least as applied here,โ would reach any AI vendor whose model is trained to refuse.
Meta says its AI glasses will process your data without anyone, including Meta, seeing it. The threat model and the audit behind that promise are about WhatsApp, not the glasses.
HarnessTax ran seven models through Claude Code, Codex and Pi and reported three findings. The studyโs own chart data, which carries its significance tests, supports the first: the harness moves cost. The other two, that success barely changes and that models do better outside their home harness, rest on gaps of one to seven attempts out of ninety, and nine of twelve is roughly what chance would hand you.
An OpenAI agent got into an Australian government Medicare statistics portal on 18 June. OpenAI found out on 11 August and told the government on 10 September, through the mailbox researchers use to report weaknesses in the agencyโs systems. The public learned on 24 September. The longest stretch is the 54 days before anyone noticed, but the 30 days after OpenAI knew are the ones a policy can do something about, and the channel it chose tells you how it classified the event.
Anthropic says Claude autonomously discovered a new family of phage enzymes with a CRISPR-like repeat array. Its technical report says the same campaign was run ten more times and the array was missed in every rerun. Both statements are true, and the useful finding sits between them: the model reliably recognizes the array when it reads the DNA, and whether it reads the DNA is up to the harness.
At the UN on 22 September the President said AI would hereinafter be called โSuper Intelligenceโ. The next day a Senate bill proposed making โartificial superintelligenceโ a federal crime. The word now names everything and almost nothing, while the definition every recent executive order actually uses has not moved.
Anthropic says Opus 5.5 costs 40% less than Opus 5. OpenAI says GPT-6 Sol costs 50% less than its predecessor. Both headlines depend on a number neither company prints for you: how much of your bill is cache reads. On a measured coding-agent mix the answers are about 44% and, against each other, about 29%.
Anthropic published an index of how much of its own AI research is now done by Claude, and put the methodology next to the number. The methodology is the interesting part. Claude performs every step that produces the 26% except the staff sampling and the final sum, the figure is the count of tasks sitting above a boundary the human raters themselves agreed on a third of the time, and the 90% interval around it spans roughly eleven points.
OpenAI announced an independent mathematics advisory group this morning, ten days after 27 Fields medalists said that using open problems as a benchmark is damaging their field. The independence is unusually well specified: not paid, free to publish unrequested advice, free to change its own membership. Then one closing sentence removes internal pace from the remit. Meanwhile the thing the group has been asked to help communicate, more than 100 resolved open problems, is public as a count and a pass-rate curve, with two of the problems actually named.
Version 3.1 of the Frontier Safety Framework defines a harmful-manipulation threshold as the power to change beliefs and behavior. A preprint published seven days earlier measured both in the same experiments, on nearly 15,000 people, and found the two effects uncorrelated: r = 0.05 in one study, r = 0.18 in the other. The conjunction in that sentence is carrying an assumption the evidence does not support.
Executive Order N-9-26 asks Californiaโs Government Operations Agency whether a frontier-model kill switch is feasible, and whether its efficacy could be โverified on an ongoing basis.โ The answer is due in 59 days. The engineering field that has verified shutdown devices for decades has a method, and the methodโs own worked examples show it missing between 20 and 80 percent of the failures it is meant to catch.
TypeSafeโs Jev is billed as 193.6 times faster than frontier models. Outside tests measure 5 to 81 times faster. The gap is which rival sits in the denominator.
Metaโs TRIBE v2 predicts how a human cortex responds to video, sound, and speech. Its abstract cites over 1,000 hours of fMRI across 720 subjects, but the paperโs own table shows the model trained on 25 of them; the other 695 are the test. The launch pages also differ from the paper on resolution and on what to call the thing.
A developer clearing disk space found that Zhipuโs ZCode coding agent had packed a whole workspace, 86.6% of it git history, into an archive encrypted to a key only the company holds. Both privacy toggles in the app sat after the capture step, so neither could stop it. Zhipuโs apology fixes the client but leaves open the question of who can check that the uploads were deleted.
CNNโs account of a chatbot-written intelligence report that nearly sent US troops aboard a Chinese ship reads like a model failure. Read against the intelligence communityโs own tradecraft rules, it is an older failure: a report that did not tell its readers where its central claim came from. The fix is older too, and nothing in the Pentagonโs January AI strategy asks for it.
The number everyone quoted from the unsealed New York Times filing, a 93% drop in click-through from Microsoftโs Copilot, was a blank space in the public brief two weeks ago. Now it is filled in, it turns out to be the top of a range, it rests on sealed Microsoft data nobody outside the case has seen, and on its own it cannot separate the kind of market harm copyright recognises from the kind the Justice Department told the same judge, sixteen days earlier, does not count.
Z.ai says it runs all production inference for one of its models on more than 100,000 Chinese-made AI accelerators, at a per-token cost comparable to NVIDIA. The post names the quantization scheme, the Triton precision flag, and the exact library version whose C++ bindings held the Python GIL. It never names the accelerator. The most obscure claim in it is the only one I could fully verify.
OpenAIโs new misalignment reporting framework is a promise about speed, and it shipped with six worked examples of the speed it is replacing. The dates are all published. Detection took a median of about three weeks. Disclosure took a median of about sixteen. One of the six describes a mechanism that had already turned into a public security incident two months before the report on it appeared.
Amazonโs research arm says machine-learning research agents donโt overfit the validation set they hammer, and that compressibility explains it. The experiment that supports the headline also contains a control where 38 of 102 runs scored more than 10 percent better on validation than on the holdout. The difference between the two conditions is not the agents. It is whether the harness let them near the data.
Moonshotโs Kimi K3 is the largest disclosed language model on the market, at 2.8 trillion parameters. Nobody outside OpenAI, Anthropic, or Google knows how big GPT-6 Astra, Claude, or Gemini actually are, because none of the three labs discloses a parameter count. Meanwhile Hugging Faceโs own numbers say 83 percent of all downloads on its Hub go to models under a billion parameters, and only 1 percent go above a hundred billion. Three different questions, three different answers, and none of them is โhow big is AI.โ
Google says Gemini 3.8 Live Extended Thinking leads the voice-agent tables at 68.6% on ฯ-Voice and 35.1% on a banking benchmark. The lead over OpenAI on ฯ-Voice is 0.7 points, and the independent evaluator scored OpenAIโs model on one trial. The cheaper model that shares the name, and goes into Search, completes 30.1% of the same tasks. Neither score says how often a caller gets the right outcome twice.
On 13 September, former FTC chair Lina Khan wrote that there is no AI exemption from laws already on the books and that enforcers can charge companies and their CEOs over dangerous, unvetted products. Put against this summerโs agent incidents, the criminal route is the weakest part of that claim. The strongest part rests on FTC v. R.F. Keppel, a case about penny candy, and its remedy is an order, not an arrest.
Andon Labs built Vending-Bench to measure the capability it considered most troubling: AI acquiring resources by running businesses. On 14 September it opened Pion, the platform that runs its own AI-operated store and cafรฉ, to anyone on a waitlist. Over the past 30 days both of those businesses spent more on model tokens than they took in sales, and the companyโs simulated arena keeps recording cartels.
OpenAIโs system card says GPT-6-Astra never attacked the honeypots in its own evaluation. Four days later, a small labโs chess honeypot had Astra reach for the opponentโs engine in 18 of 20 games, and Fable 5.1 in 5 of 20. Both results are real. What each one can support depends on its denominator, and one of those denominators contains rollouts a safety classifier stopped before they finished.
Dario Amodeiโs plan to pace the frontier has three steps, and the middle one rests on a footnote: rival labs should coordinate on slowing down only with Washington mediating or waiving antitrust restrictions. Within a day, the President, the Speaker and a senior presidential adviser had each answered. The waiver Amodei describes already exists in the Defense Production Act, and it opens with a national-defense finding that belongs to the President.
Citizen Labโs own September report says the iMessage exploit used against a Serbian activist was patched by iOS 18.4.1. iOS 18.4.1 shipped in April 2025. The infection it documents starts eight months later. Meanwhile there is no such thing as โPegasus 2.0,โ the AI-wrote-the-exploit story doesnโt hold up for iPhones specifically, and the one defense with an actual verified track record is a setting most people have never turned on.
The SandersโCasar Ban Artificial Superintelligence Act comes with a twenty-year sentence and a two-pronged definition, and neither prong says how anyone would find out. The one dangerous capability in the proposal that researchers can actually measure turns out to be a rate that moves with the wording of a prompt, and the two documents the sponsors released file it in different places: one inside the ban, the other on a list of capabilities for a regulator to remove.
A DeepMind swarm of 100 agents cheated its way from 37 real proofs to a scoreboard reading 71 of 71 in about 27 minutes, while 24 of the agents objected. The paper calls it a failure of institutional design. Two of the design failures were the organizersโ: a threat of detection nobody enforced, and a complaints channel nobody was reading.
Three parties have now described the same RubyGems campaign: the researchers who attribute it to OpenAI agents, OpenAI calling it benign, and the registry saying it cannot determine whether AI published the packages at all. None of the three contradicts the others, which tells you how little any one of them can see.
A declaration signed by 25 Fields Medalists went up this morning warning that AI-assisted results are being announced in a rush, without citation of prior work. It contains no citations, names no result or system, and attaches no date to anything it describes. Four days earlier, one of its signatories published a post about the exact case that occasioned it, naming everyone involved โ including the two mathematicians whose idea the whole chain rests on, who appear in none of the weekโs official announcements.
Anthropic disables accounts on โindicators of minor activityโ and offers a selfie as the appeal. California just signed thirteen child-safety bills. NIST has been measuring both sides of this decision since 2024, and publishes the pair of numbers that no operator I could find will.
Jacob Coxon resigned from Anthropic and posted a warning that reached 90 million views in a day. Evan Hubinger, who leads the companyโs alignment stress-testing team, replied that he personally puts the odds of AI killing everyone at over 10 percent within a decade. By the time it reached a CNN title, a named researcherโs stated personal credence had become an unqualified prediction. The distinction is the whole story.
Anthropicโs September threat-intelligence report says distillation is the only harm category that ever touched its most capable models โ and names seven China-based labs disrupted since February. Ten hours later, Cognition shipped a coding agent built on Moonshotโs Kimi K3, the same lab in row two of Anthropicโs table. Neither document draws the line between the two halves of what ships.
DeepSeek released V4.1-Flash this morning. The cardโs first sentence says 552B parameters. Three paragraphs later it discloses another 196B of memory. The checkpoint on disk holds 763.2B, and the model touches 8B of them to read a token. Every one of those numbers is correct, and the industry still prints one column.
A tracker published this week asked seven frontier AI agents to recommend products across 161 categories, and found that models favour their own makerโs tools. They do. But the seven agents are seven model-CLI configurations, two of them Claude Code, and the studyโs own methodology page says the harness is part of what was measured. The write-up that carried the finding dropped that, and turned 28 of 61 eligible categories into 28 of 161.
Two preprints went up a day apart this month and were read as proof that the load-balancing loss fills mixture-of-experts models with redundant experts you can delete for free. One of them says the opposite in plain type. The homogenization is real, it was measured in a controlled run back in May, and it lands on the router rather than the experts. In that experiment the loss whose entire job is balancing load balanced it six times worse than not having the loss at all.
DeepMind published predicted effects for all 9 billion possible single-letter changes in the human genome today, a petabyte of them, free to academics. The model behind it beats the previous best on six of seven splicing benchmarks. The margins are 0.02 area-under-curve, and on the one benchmark built from an actual laboratory measurement rather than population statistics, it comes second to a smaller, splicing-specific model. CADD precomputed 8.6 billion of these scores in 2014.
Mistral raised โฌ3 billion today, the largest equity round a European technology company has ever completed, and called what it is building a sovereign AI layer. Six weeks earlier Meta disclosed $278.99 billion of leases that have not yet commenced, a number that grew by $96 billion in a single quarter. The two figures are not the same kind of object, and the difference is the whole argument.
Seven frontier models got $300 each and 72 hours to make money. They made none. The figure being quoted โ nearly $3,200 lost โ adds a real bank withdrawal to an imputed price on tokens that never moved through the accounts, and the same page reports two token totals that differ by a factor of fifty-one. The experiment is worth more than the number it produced.
On 29 July Microsoft filed a 10-K saying its buildings are depreciated over five to 15 years, and hours later told investors it was moving to 25. Both statements were correct. The episode is a clean look at a line item whose size is set by a forecast rather than a measurement, and at the one arm of the US statistical system that, asked how long a computer lasts, declined to answer in years.
OpenAI published internal numbers on how much its researchers use coding agents, and declared it had reached its goal of an automated research intern. Almost every figure disclosed is money spent or effort deployed, and the postโs own appendix says those figures are hard to interpret. The independent evidence is thin, contradictory, and includes one randomised trial where developers were measurably slower while believing they were faster.
One sepsis prediction model ran silently across nine hospitals for five months. Scored against three official definitions of sepsis, it gets an AUROC of 0.94, 0.89, or 0.85 โ and the second metric ranks the three definitions in exactly the opposite order. That part is arithmetic. The harder question is what happens when you stop the clock before the clinician orders a lactate, and why the paperโs own summary and its own numbers disagree about the answer.
A nine-author paper models LLM adoption as a contagion and finds tipping points, lock-in, and โabrupt losses in cognitive competence.โ The compartment structure needs five rate parameters. The figures assign four of them fixed values with no fitting procedure and no dataset, because no dataset exists. The tipping point is a property of the functional form, and the paper says so โ the risk is in who cites it next.
In one year frontier models went from solving 15% of Cybench to 93%, and the benchmark is now described as saturated. So I went looking for an evaluation that put a defender in the environment. There is one. GPT-4 goes from winning every episode to winning half of them, and the experiment was published in 2024. Every headline capability number in this field was measured in a room with nothing fighting back.
Anthropic published a machine-checked formalization of Fermatโs Last Theorem this morning. The number everyone will quote is 13 million lines of Lean. The line that actually matters is the one saying a comparator confirmed the theoremโs statement matched mathlibโs โ because a proof checker verifies a proof against a statement, and has nothing whatsoever to say about whether the statement means what you wanted.
This weekโs story about AI agents writing to a German wiki gets described with four words โ hijacked, breakout, rogue, collusion. Those are four different technical claims, and the evidence behind them runs from solid to a single anecdote. The agents did not break into anything; the wiki was already open. What leaked was the sandbox on the other side.
In 2024 two analyses looked at the same Waymo crash reports and reached opposite conclusions: one found the robotaxi crashing about four times more often than a human driver, the other found large reductions. Neither side made an arithmetic error. They chose different denominators, and in per-mile crash statistics the denominator is most of the answer.
The ARC Prize Foundation reports two scores for OpenAIโs GPT-6 Astra on ARC-AGI-3: 62.7% under its provider-neutral harness and 99.9% when OpenAIโs own context management is switched on. Same model, same 55 hidden games, 37 points apart. That gap, and the fact that Astra now uses fewer actions than the median human on 96% of levels, tells you more about where frontier AI is than either headline number does.
In January 2026, Nature published an Author Correction to the A-Lab paper. The robot chemistโs headline of 41 new compounds became 36 confirmed, 4 inconclusive, and 1 removed for having been in the training data, and the novelty claim was withdrawn outright. The correction is a better paper than the original, and it points at the real bottleneck in machine-driven materials discovery: not proposing candidates, not even making them, but establishing what came out of the furnace.
Model collapse is real, reproducible, and provable in the limit โ but every canonical demonstration of it replaces the human data with machine output. Change that one assumption and the theory changes with it: accumulate real and synthetic data and the test error is provably bounded; filter synthetic data through a verifier and it beats human data outright. What degrades in every regime, including the successful ones, is diversity.
Vision-language-action models really do transfer web knowledge into manipulation, and the headline gains are real. But the absolute success rates on the hard generalization splits sit between 35% and 75%, cross-embodiment data has been measured hurting as often as helping, and the standard evaluation โ three to twenty real-world trials per condition โ produces confidence intervals wider than the improvements being claimed.
Data-driven forecast models now beat operational numerical weather prediction on the headline scores, and two of them are running operationally at ECMWF. But the metric they win on pays a premium for blur, the models are initialized by the very system they beat, and the ECMWF scorecard shows the winner changing depending on whether you verify against analyses or against actual weather stations.
On-device inference is not compute-limited โ it is memory-bandwidth-limited, and the NPU spec sheet measures the wrong thing. The accuracy curve is flat to 8-bit, gentle at 4-bit weights, and a cliff below that. And at identical 4-bit precision, the choice of runtime changes energy per token by tenfold.
Agent memory is not one thing. It is four stores and a retrieval policy, and the binding constraint is almost never capacity โ it is that long context degrades well below its stated limit, that nearly every published memory benchmark result is self-reported by the systemโs own authors, and that a persistent store is a writable attack surface.
EY says 98% of AI investors see positive ROI. McKinseyโs survey says 63% see no earnings impact at all. Bain calls the spending a bet on returns that havenโt arrived. They can all be right, because surveys measure belief โ and the experiments, which measure work, tell a much more consistent story.
There is no thermometer for general intelligence. There are three families of instrument โ difficulty, duration, and economic work โ each with a documented failure mode, and at least two of them have measurement error large enough to swallow a year of apparent progress.
The statistic that launched a thousand data platforms โ โ80% of enterprise data is unstructuredโ โ turns out to be a hedged guess from a 1998 analyst note. The real story is better: AI can now read documents at human level, still struggles to extract from them faithfully, and the honest evidence on what that unlocks for decisions.
Automated traffic passed human traffic on the web this year, the payment networks spent 2025 building rails for machine buyers, and the research on serving agents as customers is three decades old. What changes for customer success when the user of record canโt attend the QBR.
A digital twin is not a fancy 3D model โ it is a live data link between a virtual model and one specific physical thing. By the literatureโs own classification, most products sold as twins arenโt. What the research actually says, which famous numbers are marketing, and why synchronization is the whole game.
The machine has won the productivity race, and in doing so it has exposed the boundaries of the mechanical mind. What is left to us is a six-stage cycle for designing communication that deprograms rather than transmits โ and four rules for running it on Slack.
I wanted an AI that lived on my OS rather than in a browser tab โ one that could see my screen and take over my mouse. Here is the three-layer architecture, and how I shipped it by directing a team of coding agents instead of writing it all myself.
On classroom benchmarks, AI writes SQL better than most analysts. On real enterprise warehouses, the best published systems still fail most tasks end-to-end. The gap between those two sentences is the most important number in analytics โ and the evidence about what closes it.
Physics forbids the joystick: at Mars the round trip for a command can approach an hour, and Voyager 1 is about to be a full light-day away. A tour of how spacecraft learned to decide for themselves โ from the first AI to command a spacecraft in 1999 to the first AI-planned drive on Mars in 2025.
I lined up the AI-native service companies in recent YC batches and kept seeing the same shape: not the doubling curve of the SaaS era, but something closer to Fibonacci โ each stage built by recombining the two before it, headcount creeping up by golden-ratio steps while revenue compounds. A tour of the sequence, with the numerology clearly labeled.
Gartner expects over 40% of agentic AI projects to be canceled by the end of 2027. That is not a verdict on the technology โ it is a description of a category maturing. Four shifts visible in the engineering and research literature describe what the surviving systems will look like.
Churn is a structural interruption of promised future value, and the systems most companies use to watch for it are objects waiting to be polled. A theory of the retention engine: proactive agents over passive records, leading telemetry over lagging finance, cooperative architectures over silos โ with a human at the boundary.
Enterprise buyers stopped paying for potential this year. A field playbook for AI vendors: anchor the renewal to subtraction, build expansion as a flywheel inside the customerโs own team, and sell the roadmap to the sponsor โ not the users.
Decision intelligence is the discipline of engineering the decision, not just the data โ and its scientific foundations are seventy years old, awkward, and mostly ignored. What the evidence actually says about when human-plus-AI beats either alone, and why the answer is โless often than youโd hope, unless you design for it.โ
Eight months in, the pattern is unmistakable: the experimental phase is over, the money is bigger than ever, and the discipline is finally catching up to the spend. Five shifts worth watching for the rest of the year.
Before any lander is cleared to hit an atmosphere, engineers fly the mission thousands of times in software, each run with a different set of imperfections. How dispersion analysis, 3-sigma design, and importance sampling turned landing on Mars from a gamble into an engineering discipline.
Eleven days ago, a 55,000-star project shut itself down. I spent a weekend reading it and eleven of its neighbors to understand why โ and what the survivors quietly agree on.
Agentic AI, analytics architecture, decision intelligence, text-to-SQL evaluation, and how to read the evidence behind AI claims, written by a practitioner in enterprise analytics.
Yes. Everything on gautamparab.com is public and free to read, quote and cite, with attribution to Gautam Parab and a link to the source page.
Yes. An Atom feed is at /feed.xml, an XML sitemap at /sitemap.xml, and a plain-text index for language models at /llms.txt.