Solving hard problems with tech. ex-Senior Advisor to central bank, ex-IBM Distinguished Engineer & ex-CTO. Opinions expressed are own
Internet
Joined September 2007
- Tweets3.1K
- Following3.7K
- Followers2.4K
- Likes1.3K
Anthony Butler retweeted
Introducing Samaya Sterling, our new model post-trained for challenging financial research that outperforms the frontier! Sterling started off as a two person research effort to see if we could increase efficiency in agentic orchestration. To our surprise we were able to post-train a much smaller model (Nemotron 3.5 Lightning, a 30B mixture-of-experts model with 3B active parameters) to greater accuracy (58.9% vs 55.9% on FrontierFinance) and 11x lower cost ($0.61 vs $6.89) than Claude Fable 5.1, the best performing frontier model.
We did this via a four-stage training process, two stages of SFT where the model first learns to research with our financial tools and how to construct a good agentic research strategy. The second two stages are RL, with an offline GRPO stage where the model further hones its research strategy from eight teachers that provide diverse approaches, and an online GRPO stage where the model can make end to end attempts and learn from its mistakes.
We see substantial gains from the co-adaptation of the model and harness through post-training, both through the model learning to effectively use the high-quality financial tools, and better compaction approaches that let even a small model execute on longer and deeper research.
We are exploring multiple follow-up directions, from multi-agent research swarms using this small model, to training firm-specific models to encode unique judgement, to training a larger model, where early results show even greater gains from the training process.
The full research is linked below!
Anthony Butler retweeted
How on earth did this happen:
The Indian certificate authority (CA) eMudhra had one job (well, two):
1. Create secure, 384-bit private keys
2. Store these securely and do not leak them
Well, eMudhra chose private keys between 1 and 256 (the numbers!!) and called it a day. WTF
Anthony Butler retweeted
12 Oct 2026 | Grand Mufti of the Kingdom of Saudi Arabia🇸🇦:
The terrorist Houthi militia’s targeting of holy sites and visitors to the Two Holy Mosques is a grave crime with dire consequences
The Grand Mufti of the Kingdom of Saudi Arabia, Chairman of the Council of Senior Scholars and the General Presidency for Scholarly Research and Ifta, Sheikh Dr. Saleh bin Fawzan Al-Fawzan, said:
“The terrorist Houthi militia’s targeting of civilian facilities and the taking of innocent lives including those travelling to the Grand Mosque in Makkah and the Prophet’s Mosque in Madinah are among the gravest crimes, for which Almighty God has threatened a painful punishment.”
God Almighty says: “And whoever intends wrongdoing therein, We will make him taste a painful punishment.” The Messenger of God, peace and blessings be upon him, said of Madinah: “Whoever intends harm against its people, God will melt him as salt dissolves in water.”
He added: “Throughout history, no one who has followed the path of this criminal gang in targeting holy sites and those travelling to His Sacred House has escaped destruction and a grievous end. We ask Almighty God to bring these criminals to an end and to protect our country and the countries of the Muslims.”
Anthony Butler retweeted
اللهم عليك بكل من أراد أمننا وبلادنا وقيادتنا بسوء، اللهم رد كيده عليه، واجعله عبرة لغيره.
﴿ رَبِّ اجعَل هَذَا البَلد آمِناً ﴾ 🇸🇦
Anthony Butler retweeted
اللهم في هذا اليوم المبارك
نسألك نصر جنودنا الأبطال
المدافعين عن أمن بلادنا ومقدراته
بعزم وقوة وهمة عالية..
اللهم تقبل شهداءنا وارحمهم..
واشفِ المصابين ومدهم بالشفاء العاجل
اللهم احفظ بلادنا من كل مكروه 🇸🇦🙏🏼
Anthony Butler retweeted
رحم الله شهداء الوطن وعزائي لكل العائلات الكريمه ولكل السعوديين والمقيمين على ارض بلدي الطاهره 🇸🇦
Anthony Butler retweeted
In my career, compilers have always generated better code that I ever could have. I find this to be increasingly so with contemporary AIs.
But you know what neither can do better than me?
Knowing what software to write, what shape it should take, and what I want it to do.
The entire history of software engineering has been one of rising levels of abstraction.
This is how it was, how it is now, and how it shall be in the fullness of time.
Anthony Butler retweeted
I am done with this shit. It is over. The state of engineering right now is horrible. It has been half a month since I started a new role at a big company. Nobody knows anything here. The specs, code, tests, PRDs, tickets, resolution of those tickets, reports, etc., everything is made by Claude Code. Nobody on my team likes this. They are being forced to ship as much as they can. I have heard multiple times from higher management that pushing code is not a bottleneck, so why are we slow? People are working 12 to 13 hours a day just to press enter. Nobody is reading anything. Humans in corporate are doing nothing on their own. Everyone, literally everyone, from an L1 to an L7 engineer here is doing the same thing. Talk to Claude. There is no sense of victory. Nobody is resolving bugs. In reality, nobody is thinking anymore. Everything is done by LLMs. It is so soul-sucking. I would not mind it, to be honest, if we were at least given the time to check out the code and see what is going where. But no, the goal is to just ship. No matter what happens.
Anthony Butler retweeted
تذكير مهم...
التصوير والنشر يخدم العدو.
لا تصور
لا تنشر
لا تشارك أي تفاصيل حساسة
أمن الوطن مسؤوليتنا جميعاً
Anthony Butler retweeted
The original Amiga demoscene coming to life as live backgrounds in Omarchy is so damn beautiful. I watched this demo Spaceballs demo when I was 14, and my mind was blow.
AI is adopted fastest not where models perform best, but where their output is cheapest to check.
That one sentence explains most of the adoption map. New essay on why.
In 2016 Hinton told people to stop training radiologists.
2026: radiology has more FDA-cleared AI than any specialty, roughly 3/4 of all cleared medical AI. Radiologist demand held up.
The AI finds the lesion. A radiologist verifies and signs.
That is the pattern everywhere.
AI collapsed the cost of producing plausible work. It did nothing to the cost of trusting it.
Those used to be bundled. The skilled human who made something also certified it, tacitly, for free. AI unbundles them.
So why did coding go first?
Not because models are best at code. Because code arrives wrapped in the deepest stack of cheap automatic verification ever built.
Compiler. Types. Linters. Tests. CI. Runtime.
Errors caught in milliseconds, long before anyone expensive
A credit decision has none of that.
No compiler. No test suite. The first check is also the most expensive one: a qualified human reading the whole thing.
Same models. Same capability. Wildly different cost of trust, and therefore wildly different adoption.
The strategic consequence.
Models are commoditising. What a frontier lab ships this quarter reaches your competitor soon after.
The machinery for judging whether output can be trusted does not commoditise, because it encodes your firm's policy and risk appetite.
Rent the intelligence. Own the judgment.
Full essay, with the economics (Spence, Polanyi, Holmstrom, Kremer) and the predictions it commits to:
modelsandmarkets.com/p/the-v…
Companies have driven the cost of producing work to almost zero, and most of them are seeing no matching gain. More output, no more done. The bottleneck didn't disappear, it moved.
Here's the economics of why: when one input to a production process becomes cheap and abundant, value migrates to whichever complementary inputs stay scarce. AI made the input of plausible cognitive output nearly free. However, it did nothing for the other inputs: specifying what should be built, verifying the result is correct, and standing behind it. Those didn't get cheaper. One of them, the cost of catching a confident, plausible error, arguably got more expensive (as anyone who has had to deal with reading/reviewing/responding to "workslop" knows).
So the value moved to these inputs. I call the return on those scarce complementary factors the Coordination Premium.
AI has landed first and hardest in software, and the usual explanation is that models are good at code. The real reason, in my view, is that code is uniquely cheap to check: compilers, types, tests, CI, linters, and a runtime that either works or doesn't. A legal memo has no such machinery, so the first line of verification is also the most expensive one. i.e. an expert reading the whole contract or law or regulation. Same model, same capability, wildly different cost of trust, and therefore wildly different adoption. As such, I believe AI diffuses in order of verification cost, not model quality.
This then points at where the next moat sits. Models are commoditising and anything a frontier lab gives you this quarter reaches your competitor next quarter. What doesn't commoditise as easily is the firm-specific machinery for judging output of the model. In other words, rent the intelligence but own the judgement.
This new essay works the whole argument through, with Coase, Baumol, Akerlof and one very instructive train crash.
modelsandmarkets.com/p/the-c…