Finding documents for an AI agent is cheap; reading them kills latency, token budgets, and context windows. I wrote about how 1970s math gives agents a clear policy for what to open first, how to read in stages, and when to stop. Let me know your thoughts! https://t.co/1ncaoWiZeu
Now in beta: Businesses in ChatGPT.
Claim your business, manage your information, and get insights on how people engage with you.
Eligible businesses can also build a Business Agent to answer customers’ questions and help them explore what you offer.
https://t.co/N2GEZX0mFT
When Anthropic release a new model it's always worth taking a look at their updated prompting guide - here's an entertaining tip to make Fable 5.2 less likely to be cautious in how it improves existing code: tell it "This codebase is yours" https://t.co/mKjdw4QPxV
While Tech companies have been debating the future of models, consulting firms have remained unusually quiet. My view on where consulting goes from here. Stop renting the model layer, start funding it.
The AI stack is getting integrated from both ends at once. Labs and hyperscalers are pushing down into chips, compute and power to control cost, and pushing up into workflows and deployment to capture revenue. Consulting sits right at the top of that stack - application plus proprietary intelligence. Client context, transformation history, the judgement to actually move an organisation. That is the layer being climbed into.
So far the industry response has been to partner. OpenAI and Anthropic alliances, Microsoft’s $2.5B Frontier Company, AWS putting $1B into embedded engineers. Sign up, get early access, hope it stays symbiotic.
To me that is backwards. Every one of those engagements teaches the platform how enterprise work actually happens. You are handing over the one layer you own and getting tooling back.
Cursor is the warning here. Owned the workflow, rented the model, built their own to get out, still got squeezed from above and below and sold for $60B. It did not fail exactly, just stopped being independent.
The alternative is not building a frontier lab. Nobody in consulting can fund that and nobody should try. It is making model dependence non-strategic - fund the open weight ecosystem, post-train on your own data, keep the base swappable. Own the adapters, the harness and knowledge graphs, let the model underneath be a commodity you can change out in a quarter.
Thinking Machines put out Inkling in July, which is the biggest American open model so far at 975B parameters. Nemotron 3 Ultra from Nvidia sits behind it at 550B. Fifty companies just signed a letter asking Washington to keep open weights legal, OpenAI and Google among them. But permission is policy, not capital. Somebody still has to build the bench.
Open still trails on long horizon agentic execution, which is the exact capability that eats delivery work. Good enough for analysis, not yet good enough to run delivery on.
And most of the open ecosystem is Chinese. Cursor’s own escape model was built on Kimi. Trading American platform lock-in for Chinese base model dependence is not sovereignty. Someone has to fund a western open bench, and consulting firms are one of the few groups with both the capital and the reason.
None of this saves the pyramid. Delivery headcount is going either way. It is about who owns what is left, and whose models it runs on. Every platform has now said, publicly and at length, what it thinks the future looks like. The firms with the most to lose have said nothing.
I fine-tuned Gemma 4 E4B (4B params) into a buy-side due diligence analyst. Entire pipeline on one RTX 5090:
→ SFT: LoRA r16, bf16, 855 examples from SEC filings, prompt-masked
→ GRPO: fully programmatic reward (set-F1 vs gold + format bonus) — no reward model, no preference pairs
→ Splits grouped by company, so eval = unseen businesses
Zero-shot on 209 held-out cases: 0.993 F1, 99% exact match, malformed outputs 65.6% → 0.0%.
The benefits stack: house conventions live in the weights instead of costing ~1,300 extra prompt tokens per call, outputs are parser-stable,
inference is local - confidential deal docs never touch an API.
Bonus finding: during RL the model started disagreeing with my training labels on lease accounting. It had a point.
Model (GGUF/Ollama) on Huggingface 👉 https://t.co/quqRu1DEZY
Your top performers run on anxiety. Your AI runs on confidence. That's the problem.
The night before shipping, your best people aren't sleeping. They're asking: "What did I miss? Do these numbers tie out? What's a hostile reader going to attack first?"
That scrutiny is what separates good work from bulletproof work. Yet we let AI agents ship with none of it — just blind, hallucinated confidence.
So I built Anxiety Check 😰 — a skill that stress-tests knowledge work (reports, financial analysis, etc.) before it goes out the door.
Most tools check whether a doc is well-formed. This one asks whether the work is true, complete, and defensible — in three passes:
🔍 Coverage — what got missed (unreferenced sources, thin sections)
🔢 Correctness — re-derives margins/CAGRs/build-ups with real compute, catches table-vs-text mismatches and mixed currencies
🛡️ Adversarial — deferred analysis, unbacked superlatives, uncited numbers
It's diagnostic only — it flags issues, never silently rewrites your work — and won't report "clean" until the reasoning checks are actually done.
The kicker: zero dependencies so it runs in any locked-down sandbox and is harness agnostic. No pip install, no network.
Give your agents a healthy dose of professional anxiety.
Link to the repo in the comment below.
@OpenAI - why change something that was working great? Your new deep research is considerably lower quality than old deep research / legacy research. Your new deep research is right now producing research considerably lower quality than MS Copilot
What if you could talk to your diligence report? 🗣️
Introducing Coconut Jazz, a prototype I’m building that turns raw data into board-ready infographics.
I just added Voice Interaction (the reasoning capabilities of Gemini models are seriously impressive
🔊 Sound ON for the demo.
Tired of "AI" in Excel that just autofills? 😴
I built an Excel agentic AI copilot that actually works.
It uses @GoogleDevs Gemini + @langchain's LangGraph to turn complex goals into multi-step plans... and then executes them.
Its fully open-source.
https://t.co/3xNUIuihHO