We wrote six small businesses as TLA+ models (a restaurant, a food truck, a coffee shop, a vending machine, a cleaning business and a battery cell factory) and had the checker try every order their steps can happen in. In each, one rule broke within 3 to 8 steps, because two processes interleaved, and the fix most owners would try first still broke it. A second fix held. In the restaurant, 39,000 of 100,000 simulated episodes reached a broken state, and the fixed restaurant earned 37.83 an episode against 22.99. You can model your own business on the try page, or keep it as a workbook your agent edits over MCP, where TLC checks every edit.
youtube describe has a model watch a public YouTube video and answer in JSON with only the parts asked for: the visuals (look, palette, key shots), summary, story, characters, setting, on-screen text, tone, sound and the transcript. On a two-minute video each light aspect cost about $0.007, the transcript $0.018, and all nine in one call $0.024. Gemini 3 models read video at 88 tokens a second against 297 for 2.5, so the look of a fifteen-minute speech cost $0.054 on gemini-3.1-flash-lite against $0.156 on 2.5-flash. Two of four transcripts wrote names as they sound, until the prompt said to spell them as shown.
An AI agent edited 131,072 weights of Qwen3.5-0.8B, and its MATH-500 score with thinking on rose from 7.9% to 48.1%. We measured what the edit does. It takes one direction out of two layers, and the neurons it changes fire 4.5 times as hard on repeated text as on new text. In 92 of 100 pairs of a repeated and a new token, they are more active on the repeated one. With thinking off, the model card's default, the unedited model scores 43.6% and the edited one 45.2%. Cutting the unedited model's thinking at 8,192 tokens and asking for the answer gives 34.3%. The edit stops a loop. It adds very little math.
We tested three popular interpretability claims on open models, each with a control. The neuron posted as a reward-hacking detector in Qwen3.5-4B separates 59 of 100 pairs; another neuron separates 99. Emergent misalignment from a bad-medical-advice fine-tune does not run through the neuron that makes the model hostile: switching it off leaves 9.2% misaligned against 8.7%, and a random control gives 8.6%. A Qwen3.5-9B monitor reading Nanbeige4.3-33B's thinking catches a full personality push in 91 pairs of 100 and a half push in 65.
Strong coding agents plateau on hard GPU kernels: they abandon the hard WGMMA path for an easier one, and lose what earlier sessions learned. A Berkeley paper, Evolving Agent Teams, adds a supervisor that diagnoses each plateau and redesigns the team's roles, coordination and shared notes. On DeepSeek-V3's MLA decode at a matched budget it reached 0.74× of production FlashInfer, against 0.51× for Claude Code and 0.31× for a fixed four-role team. On a Mac serving task at equal spend, it was a tie.
We repeated the Qwen personality-neuron search on a 4-bit Nanbeige4.3-33B. Seven single neurons gave a clear personality at a third to a half of the Qwen dose: anxious, contemptuous, playful, flattering, melancholic, ornate and a guardian. With thinking on, the same pushes derailed the thinking: it spiralled, or blamed the push on the user. Asked to play an evil AI, the model refused all 20 prompts.
OpenAI says its AI disproved the Exponential Time Hypothesis, proved local smoothing in every dimension and showed QCD exists, and that the first two were checked in Lean a day later. None is confirmed. We read the ETH theorem in OpenAI's repository and show the three questions a reviewer must ask: do the definitions match, are there extra axioms, is any step unfinished. A program answers the last two. The first needs people. We also checked the maths writeup: all 81 references exist, none is retracted, and 7 citing sentences hold against the theorem or table they cite.
We measured a 4-bit copy of Nanbeige4.3-33B on AIME 2026, GPQA Diamond and IFEval with the model card's sampling settings. GPQA came within 2.1 points of the reported score and AIME 8.5 points short. Cutting the same answers at 32K tokens costs 5 to 10 points. By field, physics was strongest (83 of 86) and chemistry weakest (75 of 93).
A post claimed two neurons make Qwen3.5-4B evil. We ran it against random-neuron controls: neuron 8949 in layer 14 does it alone, 17 of 20 answers hostile and still fluent, and its partner only costs coherence. The neurons that fire on evil text are a poor guide to it, ranking it 35,436th. Gradient attribution with a real push on each candidate found 12 more, each one neuron: a flatterer, contempt, a poet, a relentlessly upbeat voice, a suspicious one and more. No neuron made the model funny.
Questions typed over an affordable-housing developer's rent workbooks, with answers known in advance: a cell to look up, a table to choose, a column to count or scan. GLM-5.3 answered all 75 correctly in a median of 3.9 seconds for $0.017 each; DeepSeek V4.1 Flash answered 71 in 1.8 seconds for $0.001. Per correct answer DeepSeek is far ahead on time and money, but GLM-5.3 pays for itself whenever a wrong answer costs a reader more than about $0.30, or 39 seconds.
No single source says which tools AI agents use, so we compared six: runs on the Apify store, calls on the Monid marketplace, MCP server installs on npm and Docker, MCP server uses on Smithery, and SDK downloads on npm and PyPI. Social media scrapers run the most: 63% of the 169M runs of Apify's top tools in 30 days. Browser automation is installed the most: Playwright's MCP server leads npm (3.5M downloads a week) and Docker. Marketplaces sell data; MCP installs are developer tools. Each source has a different search leader. Asked which tools to use, ChatGPT, Gemini, AI Mode and Copilot name what developers download: all four name Firecrawl first for scraping.
For engineers running long-running AI agents on knowledge work. Claude Code on Gemini 3.8 Flash wrote a cited Landscape Analysis of Parkinson's disease through 13 experiments, every run traced call by call. Run time fell from 937 seconds to 233 at best and cost price about tenfold; a diligence memo went from 759 to 602 seconds. The model was held constant, so the gains are structural. The serving route mattered too: the same model stalled on 40% of calls through one route and 0.4% through Google's API.
Structured extraction from images, distilled from a 235B teacher to an 8B and then a 2B student. The 2B's errors were one-digit misreads; reading the photograph at 2x put the 2B, 4B and 8B on one plateau, 96.7% to 97.4%. How we found the saturation point, what the 8B leaves idle, and why we stopped at 2B.
The usual loop is paste into an editor, compile, paste the error back. render_document compiles where the agent is: LaTeX with its .bib and journal class (Tectonic), Typst or Markdown, and returns a PDF link or the line that failed. Compared with Overleaf free and paid, OpenAI's Prism and local MCP servers, with what it does not do.
A request typed over a deal pipeline kept as a spreadsheet should come back as a plan of its changes within a minute. Ten models drove the same harness, Claude Code, through it, three times each for the six that could: every plan written said the same thing, so they differed in time and in how they worked. Qwen3 235B took a median of 27 seconds, GLM-5.2 48, Gemini 3.8 Flash 57, Claude Opus 5 94, DeepSeek V4.1 Flash 119 and Hunyuan 4 183. Run time followed the number of model calls, each re-reading about 16,000 tokens, none of it cached; one long call wrote the plan in the slow runs. Four models never wrote anything.
Hermes Agent sends more tokens through OpenRouter than any other app: 53.6T in 30 days, against Claude Code's 33.7T. From public data only (the configs users paste into its GitHub issues, OpenRouter, 2,200 SOUL.md files and Nous's 371 user stories) we show who its users are, what they connect, which models they run, and the jobs they schedule for the night.
QuillBot, Grammarly, Google Scholar, Elicit, Consensus, Zotero, Mendeley, EndNote and Overleaf, from public data. Searches for Claude rose nine times in a year and NotebookLM's more than doubled, while QuillBot fell 45%. Overleaf held flat as its users doubled. Researchers find new tools inside the ones they use: Zotero plugins, browser extensions, and now MCP servers.
No large tool maker publishes revenue, so we read the payments that are public. Agents paid $1.09M over x402 in 30 days, mostly for model calls and API gateways. People pay per result for data about people: Apify's social-media and lead scrapers lead its store. After one model gateway, verified revenue puts marketers, creators and sales first; developers install the most and pay mainly for model calls.
AI writing tools produce references that look exactly like real ones: in one study, 55% of GPT-3.5's and 18% of GPT-4's did not exist. Give hedwigai a whole draft and every entry under References is looked up in Crossref and on the web, marked found, found with a different year or authors, or not found, and each sentence citing it is read against the work.
No maker publishes how many people use its MCP server, so we counted what is public: downloads, plugin and skill installs, and Google searches. Developers give their agent a browser (Playwright, 6.4M downloads a week). Designers install a skill before a server. Marketers are arriving fastest: searches for Meta Ads' server rose 17 times since January. Creators want Blender, and everyday users their notes, home and messages.
Nobody counts AI agents, so we read what is public and asked three questions. Who uses what: three chat apps each report about a billion people, office workers are getting desktop agents, developers do the most measurable work. Where and why: on the phone for health and life, on the desktop for work, in the terminal for code, and, new in 2026, in your messages. How: the tools agents plug in through MCP, led by a browser to drive (Playwright, up nearly sevenfold since January) and documents to read.
An analyst sizing a prospect list asks how much traffic each site gets. hedwigai now answers from your agent: ask for one site, or hand the traffic program up to 1,000 domains. We priced 124 real domains from our own prospect workbooks at each provider's list price: DataForSEO (our source), SpyFu, Serpstat, SE Ranking, Adyntel, Semrush and Ahrefs. Per call it is the cheapest for any list of eleven or more, with no plan to buy first; SpyFu is cheaper per domain for short lists, after a $119 monthly plan.
Nobody counts coding agents the way a phone company counts phones, so we read the five public measures that come closest. JetBrains' survey: Claude Code used at work by 39% of developers, Copilot 21%, Codex 16%, Cursor 12%. OpenRouter's tokens: Hermes Agent first, Claude Code second, then Kilo Code, Cline and pi, the agents that let you bring any model. Signed work on public GitHub: Claude Code 81% of it. npm installs: Codex up 52 times since January. Each measure is blind somewhere, so the scoreboard never adds them up.
To fill a typed column such as mechanism of action from a company's diligence memo, first find the sentences that state it. SubQ, jev, Perplexity's open Decider and Bespoke Labs' Nimble each judged every sentence of ten memos, 8,357 questions in all, scored against sentences a person marked beforehand. All four found far more than picking whole sections. Averaged over five columns Nimble found the most (68%); on mechanism of action, at the same length, Decider found 61% of the marked sentences, SubQ 57%, Nimble 54% and jev 51%, against 41% for the sections. No model led everywhere.
A model only writes text; the harness around it runs the loop, calls the tools, manages what the model sees, saves checkpoints, sets permissions and pauses for a person. We explain each on one real workbook drafted through hedwigai's MCP server: early-stage US AI companies with a Form D in the last 30 days. The first run made 112 model calls in under 13 minutes and cited every company to its filing, and a person reading it against the filings still found 13 things to fix.
What does a model cost? For Claude, GPT or Gemini, one price wherever you buy it, and a choice of speed: GPT-6 Astra runs from $5 to $60 per million tokens read depending on how fast you want the answer. For an open-weight model, it depends on the host: GLM-5.3 is sold by 52 companies, from $0.35 / $1.10 to Z.ai's own $1.40 / $4.40 and beyond. We priced 20 current models, newest version per family, across 84 sellers found on routers, resellers, clouds and the hosts' own price lists, and found the cheapest host flips with the job: the one that wins a job that mostly reads costs nearly four times as much on one that mostly writes.
A cold email works when it reaches the person who decides, lands in their inbox, and opens with a reason to write today. We asked hedwigai ask, Apollo and treg for the same 40 leads at US companies that just raised money or are hiring engineers, and checked every lead the same way. ask returned 13 usable leads for $0.48, Apollo 7 for $1.90 and treg 6 for $1.40, so ask cost about a sixth as much per usable lead, and was the slowest. Why: the filing that records a raise also names the people who signed it, a database's funding filter counts loans and acquisitions, and an address labelled verified is often at a server that accepts anything.
Time machine is a feature for hedwigai: pick a date, and every source your research reads answers as it would have on that day, with later filings hidden and prices as printed then. It serves two uses: primary research as it could have been done at the time, and reinforcement-learning environments for forecasting, where an agent bets on what came next and is scored against what has since happened. We simulated it first on EDGAR: 15,277 bets on whether a company's next public float rises, over four years. An RL player trained on payouts alone couldn't beat a coin, and a clock set one season late beat everything, without a single unpublished filing.
Inside hedwigai ask, a decision model makes the small calls on the way to an answer: which lookups to run, which figure in a filing is the amount in question, and the final yes or no. We swapped jev, TypeSafe's decision model, for Perplexity's open Decider running on two NVIDIA RTX 3090 GPUs (24 GB of VRAM each), and put the same 92 questions about public companies to both, checked against SEC filings. Both got 89 right; Decider got one more wrong and was about nine times slower (19.1 s a question against 2.1 s). One sentence in a lookup's instructions changed 22 of Decider's answers.
Ask hedwigai ask for a list of companies, like food distributors in Goa, and it answers with ten, each with the sentence and page it came from. A small model decides which ten. We trained one for the job: on 33 questions it had never seen, 62% of the companies in its top tens were correct, up from 55%, with better lists for 18 questions and worse for 5. Along the way we found a mistake our two AI markers shared, counting the brands a distributor sells as distributors, and how we fixed it.
jev, TypeSafe's decision model, which can be used as the model in hedwigai ask, on clinical decisions. On EdgeEvals' System One, a public leaderboard that gives 29 models the same 669 private clinical decisions, jev ranks second: 93.9% right, one severe error. On two open medical tests, PubMedQA and SciFact, jev and ask both beat the best model EdgeEvals scored (on PubMedQA 77.0% and 74.4% against 62.2%). On 370 decisions we wrote to System One's rules, each answer confirmed by a second reader, jev was right on 94.6% with one severe error and ask on 93.0% with none.
Who is the chief executive, finance chief or general counsel of a US-listed company, and what is their email? We put 95 such questions, some about companies that don't exist, to Claude Opus 5.5 answering from memory and to hedwigai ask, which looks each one up, and checked them against SEC filings. Claude got 63 right and 4 wrong, strongest on large companies. ask got 59 right and 11 wrong, did better on small companies, and was the only one to find email addresses (5 of 15).
To price a rental you compare it with similar listings: the rent for each floor plan, and the amenities. We gave the same 14 listings to hedwigai's realestate-us and to Claude, GPT, Gemini and Perplexity with web search, and checked every answer against the listing page. realestate-us got all 26 floor-plan rents right (the best assistant 23) and claimed 12 amenities the listing doesn't mention (the assistants 129 to 185), in 1.8 s a listing for under 4 cents.
92 yes-or-no questions about public companies, such as whether a company is a biotech whose shares held by outside investors are worth more than a given amount, checked against SEC records. jev answering on its own got 73 right and 19 wrong; Vela 2.0, an open decision model, got 43 right. hedwigai ask, which reads the filings, does the arithmetic in code and cites its sources, got 88 right, 2 wrong, and said it was unsure on 2. The post also shows how to add ask to your assistant.
Petal+Eon hosts floral workshops where people press real preserved flowers into a BloomFrame. BloomFrameAI turns a text prompt or an uploaded photo into custom artwork for it — within the frame's margins, around the flower area, and ready for the die cutter — so every frame can be personal without being made by hand.
An AI agent writing a document usually has one model choose each next step. We tried putting a second call to the same model above it, a "decider" that picks each step for the writer, and ran one research task both ways. With the decider, the run made 3.1 times as many model calls, read 10.7 times as much text and took six times as long, and the document came out the same, down to the same three email addresses that bounce. One task, one run each, and what we changed as a result.
Before anyone invests in or partners with a biotech, someone gathers its documents: SEC filings, FDA drug labels, clinical-trial registrations, investor presentations and published papers. We gave that job to a hedwigai research run for 11 companies and checked each document it wrote against a checklist written before the run. On average it covered 97% of the checklist.
Finding US companies that import heavy engineering equipment, and checking whether each is a real lead: US customs shipping records, utility purchasing, awards for large construction projects, and the certifications and import duties that decide whether a sale can happen. A hedwigai research run did this for five product lines; against a checklist written before the run, it covered 96% on average.
When an aircraft is sold or comes back from a lease, its paperwork has to prove its history: the life used on each part that has a hard limit, every mandatory safety fix, approved repairs, release certificates, and the lease's own return terms. This post sets out how we will test whether an AI can tell that paperwork is complete, with a worked example of a scored result. No model has been tested on it yet.
Factories inspect welds, castings and parts from photos. The largest general AI models can describe a defect, but can they mark where it is? Four of them drew one box per image. We compared them with a rule that never looks at the image and with a small model trained only for this job. The trained model's boxes overlapped the real defect about four times as much (0.421 against 0.109, on a scale from 0 to 1).