Founder & CEO for @santorilabs. Ex-GM @scale_AI (human data for frontier models), ex-@Google. Writing about how AI changes how we work and think.
San Francisco, CA
Joined April 2013
- Tweets271
- Following178
- Followers204
- Likes249
here is the thing: your best people are motivated when they are surrounded by great people, in an env where they can thrive.
The rest are either empire builders or slackers. I think we still need a few empire builders. but that's it
Just spoke to the Head of AI at a huge company.
The #1 thing they're struggling with is... what do I do with the AI laggards at our company?
Like most companies, they’ve found that a Super Pareto Principle is true within their org. 10% of their employees are responsible for 90% of the uplift in productivity/token usage from software engineering to non-technical knowledge work.
They’ve spent a lot of time & resources on AI enablement from LLM foundations to function-specific trainings and skill building. And yet, the power law of productivity increases hasn’t budged.
So now they’re left with a few sobering thoughts and hard questions:
1) 90% of people are overpaid and 10% are underpaid
2) We’d be better off allocating tokens proportionally to the power law, but how do we do that without killing culture?
3) Significant layoffs are likely warranted and wouldn’t impact overall output, but cultural backlash is a big concern?
4) How do we hire more power law people as a non technology company?
5) Is all of our efforts on training and enablement ACTUALLY worth it?
6) How do we not experience a brain drain where the top 10% all leave because they want to be surrounded by other AI-native 10x’ers?
banning juniors from ai is how you get seniors in 3 years who can't tell if the output is sh*t or good
Interviewed an engineer friend who works at a traditional major bank.
Very interesting observations:
- Junior engineers are not allowed to push code with AI. Claude Code & Codex are banned from
- Senior engineers can freely use Claude Code & Codex, but only to check their and others work.
- If you push AI-generated PRs, you will get a bad rep within the organization.
- They still have a QA department that does all the reliability on their software.
- He only uses Claude for his personal projects, which there is a lack of, because he works 9 to 7.
Sincerely asking what happens to these people.
If you are not using the frontier technology, it's either that the institution gets replaced, or that eventually, these people get laid off to realize suddenly that the world has moved on, and they just stopped in the pre-Claude Code era.
What a read
Latest (and probably penultimate) Building Slack post is up: buildingslack.com/salesforce…
What it's like being eaten by the whale
we aren't talking about this enough - ur org and culture will fail it AI adoption. was already the case but now the opportunity cost cannot be higher
buildingslack.com/salesforce…
wow how is this possible if its a pattern generation machine?
well this is when you spend a lot of $$ and hours with the best experts. essentially these experts refine their own work so much that it beats some expert by default.
world is crazy
Massive. New Harvard study finds that AI-written paper reviews now beat the best human peer reviews on every measure the authors tested.
Peer review is slow and often misses errors. Earlier studies concluded AI should only assist human reviewers, not replace them.
They used 134 reviews from rapidPeer, a Harvard service that pays expert reviewers well. Claude Opus 4.8 reviewed the same 1124 papers, and GPT-5.5 plus 3 former journal editors judged the pairs blind.
GPT-5.5 picked the AI review in all 132 decided pairs, and the editors picked it 87% of the time. Only 5.5% of the AI's claims about the paper were wrong, versus 15.9% for the human reviews.
Teams that depend on expert critique can now let AI do the first full review pass, and keep a human for the final decision.
100%. For now and will be for quite some time:
1. You need to what to ask and how to initiate it. Most people still use it as a Google search
2. For complicated stuff, you need enough understanding to know how to verify it or steer it.
3. It's still a next token generator and you get to decide what it is generating on. A.k.a the question and steering you do, or the harness you built so you don't need to do that
The most important one of smart or not - can you look at the output and accurately access if it's sh*t or good?
Stop looking for an ai agent that auto-rotates API keys for you.(Or requesting a startup for that)
Instead. Paste this prompt to your agents wth computer skills and switch to federal identity instead. Don't have it do the generation part.
-----------------------+
Go through every AI/API provider we use and get rid of static stored keys where we can.
First make a list. Check the code, .env files, CI secrets and every host's env vars. For each key, write down the provider, where it's stored (every place), and what uses it. Show me the list before changing anything.
For each provider, check their current docs for workload identity federation. Don't go from memory, this changes fast. Also check whether our host can issue OIDC tokens that provider accepts. Right now I think it's OpenAI and Anthropic, plus Bedrock/Azure/Vertex through cloud IAM, but verify beyond this.
If it's supported, migrate it:
- Open the provider's console page in my normal browser with `open <url>` (or `xdg-open`). Don't use your built-in browser. I'll do the clicking.
- Tell me exactly what to create: issuer, service account, mapping or rule. Match on the exact sub, minimum permissions, short token lifetime.
- Then wait. Ask me for the IDs you need (rule id, service account id, org id etc). These aren't secrets.
- Switch the SDK to workload identity, and remove the key env var everywhere, because a leftover key can win over federation.
If it's not supported, rotate the key instead:
- Open the key page in my browser and ask me to make a new key.
- Don't have me paste it in chat. Give me a command that reads it hidden (`read -s`) and pipes it straight into the host's CLI or secret store.
Either way, staging first if we have it. Prove it works with a real call, and make sure the old key isn't being used anymore. Then tell me which old keys to revoke and I'll do it. Don't revoke anything yourself.
One provider at a time. If something breaks, stop and tell me.
At the end, give me a table: provider, what you did (migrated, rotated, skipped and why), where it changed, and which old keys still need revoking.
An urgent notice from the Railway team
blog.railway.com/p/security-…
you can really think about your P0s, or you can just mindless close your P2s.
which one is it?
the same thing that has always been true of engineering is still true. Slow down, prioritize well, do less in parallel, elevate getting things #1/2 done over maximizing utilization of your overall attention (even if it means some downtime for you)
Yuting D. retweeted
I think Duet and Dots are the wrong form factor for the enterprise because they aren't AGI-pilled enough.
The major agent releases today (Grok Bot, Duet, Dots, Muse, Instinct) all act as an "AI Chief of Staff." This allows each employee to get their work done faster, but it doesn't always make a company move faster. Executing is no longer the bottleneck, so even proactivity on individual tasks isn't necessarily a big win.
Models need to understand organizations and align with the overall mission to truly move the needle. They have demonstrated goal-directed strategic behavior in simulations: alignment.anthropic.com/2026…. The missing skill is aligning the models to the right goals. Until then, humans have to continue solving coordination.
Anecdotally, I've heard "in the high MTS circles" that a lot more time for frontier researchers is now spent on meetings for coordination since agents are doing most of the implementation and execution. The rest of the industry will soon face this bottleneck.
The AGI-pilled form factor is a model with autonomy that acts at the same level as other employees. A true collaborator rather than someone to manage.
