FDE @openai | Prev @BuildWithGrove @poktnetwork @getsoils @waymo @magicleap @twitter @google | NΨ 1T4 @UofT | opinions are my own |
Bellevue, WA
Joined May 2011
- Tweets9.5K
- Following862
- Followers2.6K
- Likes15.6K
Pinned Tweet
I keep thinking that the path to Safe AI is going to involve just as much focus on Saving Human Intelligence as it does on building Safe Super Intelligence.
Having spent the last month in “school mode,” I’ve felt firsthand how much harder it is to learn new concepts in your 30s, complete tasks for speed, and rebuilding skills that haven't been exercised in a while.
Getting used to leveraging AI as a crutch over the past few years doesn't help.
I’m increasingly convinced that Wall-E style mental atrophy is far more likely, and potentially more dangerous, than a Terminator-style Skynet.
If someone is out of physical shape, the signs are obvious and visible. If someone is out of mental shape, it's much harder to gauge.
The challenge is that being physically or mentally fit doesn't always have direct commercial ROI when things just need to get done. Economic incentives align more closely with short/mid term outputs, with long-term trends only coming after the former has been achieved.
It's kind of crazy to think that within 10 years, some derivative of Neuralink, or a product inspired by it, could have its own GLP-1 moment.
What's even more interesting is that the breakthrough may not be a pill at all.
It may simply be having the data and observability needed to precisely tune the stressors required for learning, adaptation, and growth. Just enough challenge to improve, without tipping into burnout.
This is 💎. Thinking of framing it.
I figured if I want human attention, I'd better put in human effort.
So I have adopted a rule of thumb that when I want humans to read, I hand write every single word.
I have been experimenting with this rule for a few weeks and very much liked it. For my readers, my manually written text is much more concise, so easier to digest and read. For myself, it's the process of writing that clears my thinking and gets me in to the flow zone that I was never able to get to when writing with ai.
TIL about Teacher Forcing in training.
A teacher grades a student's multi-part exam one step at a time to avoid cascading failures.
1. Student answers a step
2. Teacher grades the step
3. Teacher provide the right answer to the step
4. Student answers the next step
This forces RNNs to stay close to to a ground-truth sequence.
en.wikipedia.org/wiki/Teache…
What makes FDEs at @OpenAI special? The people.
I work with individuals who have the agency to act, and OpenAI creates an environment that gives them room to follow through.
Everyone here is as passionate about understanding the user’s problem as they are about the technology that can help solve it.
We share tools and lessons across workstreams, so everyone can build on what others have learned. The org is still young, and our collective experience is just beginning to compound!
What surprised me most since joining is the amount of founder-like energy across the board.
If you want to hear from these people directly, register for the panel!
forum.openai.com/public/even…
#OpenAIForum
👋 We heard your feedback on FDE ! We’re partnering with the OpenAI Forum on a series of livestreams, and this first one covers the basics.
Join @colintjarvis , the Head of FDE, as he leads a panel of OpenAI FDEs on how we close the deployment gap.
forum.openai.com/public/even…
Still can't make up my mind on software factories...
This presentation made it even harder.
On one hand, if you've been in the tech industry for a while, you see the need.
On the other hand, as LLMs are capable of more and more, the upside of any particular software factory seems to diminish. Customization becomes "free" and the cost is maintenance in production.
No code tools are fine, but most companies in this space struggled to scale.
Eval frameworks feel very similar. You need so many domain specific nobs and switches, that it's easier to customize it for each specific use case.
The "moat" becomes not a product that you sell, but intuition around how others have been and rebuild in a very particular way.
Curious to see if/when one of them succeeds.
It's pencils down, people. Writing code by hand is no longer an economically viable skill for most programmers at most companies. But the future of making software has never been brighter. Don't you dare black pill this beautiful moment! youtu.be/vDjW_dRyKXY?si=6Fsf…
Me on a Thursday in NYC: "I want to go see a comedy show on Saturday."
Reserves two tickets to @ComedyCellarUSA original MacDougal location.
Shows up with a friend and gets great seats.
The last two guests of the night:
- Melissa Rauch from The Big Bang Theory
- Ray Romano from Everybody Loves Raymond
This is now my new baseline for a random saturday night show.
Postgres had its moment. Sqlite is becoming the killer tool for agents.
There's a reason @supabase is hiring a supalite engineer.
🎩 💁♂️ to @chander for being early on this.
jobs.ashbyhq.com/supabase/b7…
MCP vs CLI for LLMs is the wrong discussion. People are debating at the wrong layer.
Both work incredibly well as long as they are run through a repl like environment that can persist state. The funny thing about CLI is that CLI tools are usually accessed through BASH which happens to be a repl with persistent state (the file system), therefore cli works much better than mcp. But this isn't even close to being an intrinsic property of MCP. Just means that we need better harnesses.
Right now the best repl for this are:
- bash + fs
- jupyter kernels
- codemode type repls, usually quickjs
LLMs understand the concept of forward evolving a system to solve a need very well. This comes from the agentic RL. They understand how to change the state of a codebase or acquire data from databases and then operate on it similarly to humans. But without an execution environment, they actually cannot do this properly.
My bet? Sometime soon someone will (or has already?) create a embeddable, sqlite style mini execution environment that parses bash, typescript, or tool calls into a common IL execution plan that's easy to security check before executing. And design this specifically for durable execution environments. Then we will just connect cli, mcp, webmcp, whatever to that and it accepts any of the input modalities as they can all be represented as each other.
A great short post from @simonw on rewrites.
Having been a participant, instigator and a lead of rewrites, my recommendation is to not do it.
Finding ways to compartmentalize and rewrite interfaces, APIs, libraries/packages of existing systems is much more likely to succeed.
Catching up to a moving train is much harder than fixing an existing one.
Evening view as I sip on tea, talking to my agents, through my smart glasses, kicking off hour long jobs on half a dozen remote dev boxes. It’s a lot to process.
I’ve seen a dozen quotes calling this out as a “must-read”, and I’ll ➕1 that.
Adding a few notes related to shaping the build.
1. Partnering with your user. If you’re building for yourself, do you use it every day? If you’re building for someone else, how upset would they be if the product went away? This is now the job of every engineer.
2. Momentum vs evaluation. Identifying when to focus on forward momentum VS evaluating the output of your results is a delicate balance. Evals are both an art and a science. This comes with experience and shots on net. This quote captured it well:
> “For example, knowing when to quickly build an MVP to take to users for testing, and when to slow down and take longer in order to build more carefully.”
3. Adapting your workflow. Things are changing fast. You don’t want to change your workflow every week, but you also shouldn’t be stuck in how you did things 6 months ago. You also need to know when to ignore the noise. Easier said than done. I suggest looking around at how people whose output you trust are operating day-to-day.
The world underestimates that stripe could very well be the primary backbone of all global economic activity over the next 1-2 decades.
Between wallets for agents with @link (and @privy), MPP, Tempo, MCP, Stripe CLI, Agentic Commerce Suite, sandboxes, OpenRouter, Metronome/metered billing, Radar and distillation/token fraud, it turns out that "the economic infrastructure for AI" implicates a lot. Pretty much every layer of the Stripe stack is changing, and it’s a very good time to join if you want to build high-scale software that facilitates the practical deployment of this new intelligence.
Someone asked me how we can trust AI models given that they’re non-deterministic:
We have Organic Intelligence and Artificial Intelligence. Neither is Perfect Intelligence. But, the latter is faster, more capable, and can process a lot more data, so even though they’re not perfect, they’re a lot more reliable.
How the software creation lifecycle has changed:
Before: Write. Review. Clean. Test.
Now: Generate. Understand. Sculpt. Test.
I spent four years building blockchain infrastructure.
The customers didn’t care about the blockchain.
They cared about reliability, performance, cost, and whether the thing worked.
I wrote about why you probably shouldn’t build a blockchain.
olshansky.substack.com/p/why…
