Stealth startup 🤐 | Ex Netflix, OpenAI. Kiwi, and cat dad.
San Francisco, US
Joined November 2008
- Tweets1.8K
- Following2K
- Followers1.6K
- Likes24K
Aish Fenton retweeted
Sartoria Solito is a small, family-run tailoring business based in Naples, Italy. Like most Neapolitan tailors, the family makes a very soft coat. They use only a small amount of haircloth and canvas through the body, along with a 1.5-ply shoulder pad across the shoulder line. As a result, their jackets tend to look softer and more relaxed than their London counterparts. You can see this in the photos below. Notice how gently the shoulder line falls, without the sharper structure you might associate with English tailoring.
Since Neapolitan tailoring tends to be lightly constructed, the jackets can also sit closer to the body without feeling stiff or restrictive. I think of Solito’s garments as a slightly more “modern” interpretation of classic tailoring. The cut is slim, but not as slim as much of what you’ll find in ready-to-wear. More importantly, there’s a considerable amount of shape worked into the garment, particularly through the chest and lapels, which gives the jacket a sense of dimension rather than simply making it narrow.
Consider Solito if you want a slightly more modern—perhaps even sexier—version of classic tailoring, but don’t want something as trend-driven as much of today’s ready-to-wear. Since the garments are bespoke, the house cut can also be adjusted around the margins.
Agree .. huge deal. Now hopefully we see start to see them in other harnesses too.
I have not seen near enough excitement about MCP Events developers.openai.com/plugin…
Right now, the only way most people's agents can wake up and do something is either A) cron, or B) you decide to message it
Event-driven triggers are a huge deal
Aish Fenton retweeted
Sharing some weekend reading. Faraday is a pretty interesting paper.
They train a 27B model to replicate results from AI research papers, with Codex available as a tool. The benchmark results are good, but a few things in the setup stood out more to me.
“Scientific taste” may actually be trainable. Faraday doesn’t seem to get better mainly by writing better code. It gets better at deciding what experiment to run, what evidence would actually support the claim, and when Codex has found a shortcut that produces the right-looking result for the wrong reason.
Their RL setup is also pretty interesting. There’s no clean unit test for “did good research.” They turn paper replication into something judgeable enough to train on, then use task-specific rubrics and turn-level credit assignment to learn which decisions in a long trajectory were actually useful.
It feels like a useful recipe for many long-horizon agent tasks where the reward can’t be perfectly verified, but good work can still be judged reliably enough.
The architecture is also kind of the reverse of what I expected. Most agent systems today put the biggest frontier model in the middle and have it route to tools, smaller models, etc.
Faraday’s 27B model is the researcher. Codex is a tool. The smaller model learns when to call Codex, what to ask it to do, whether the result is actually useful, and what to do next. The model with the most raw capability doesn’t have to be the one making the highest-level decisions.
This may have pretty important implications for enterprise agents too.
Maybe you don’t need to train a giant model to know everything about a company. You train a smaller model on how the company actually gets work done, what to delegate, what good looks like, which shortcuts aren’t acceptable, when to escalate. Then let it call whatever frontier model is best for execution.
Aish Fenton retweeted
We ran the largest open experiment on how frontier models do AI research.
100+ autonomous runs across 10+ models, sandboxed on 8xH200s for up to 8 days, iterating on the nanoGPT optimizer track.
Best runs closed 82% of the gap to a record built by dozens of humans over months.
Aish Fenton retweeted
Which state-tracking problems can transformers length generalize on?
Transformers' performance on different finite-state tasks varies drastically, even when the tasks are structurally similar. Now, for the first time, we can precisely predict when length generalization occurs.
Aish Fenton retweeted
Tyler Cowen: "The simple truth is that a lot of immigrants are more ambitious, more willing to endure hard work, and they have less of a stake in the established order..they are thus helping to make America more American, in terms of industriousness and creative spirit.."
Amen.
The first coding model that can actually explain itself _clearly_ is going to take a heap of marketshare.
I think ultimately Luna is going to be more impactful on the tech industry that Sol and the other mega-big models. The ability to (cost effectively) embed ambient intelligence right across a product remains under-explored.
Aish Fenton retweeted
I like the suit on this guy. You can see how a slightly longer jacket, combined with an extended shoulder line and higher rise pants, produces a more classically masculine silhouette.
Aish Fenton retweeted
I cancelled my $10/mo Calendly subscription and vibe coded my own with Fable for $12,000
Fabel is a good model, but boy-howdy are its explanations confusing at times. My fav from today: "The scope lattice is coherent and mutually exclusive" 🤔
Aish Fenton retweeted
Finally able to talk about what I've been heads-down on for 6 months at @nvidia 🦀⚡
We just open-sourced cuda-oxide — an experimental rustc backend that lets you write CUDA kernels in pure Rust.
No DSLs. No FFI. No source-to-source step. Single source.
Short🧵👇
🤖 Made with AI
ALT cuda-oxide is an experimental Rust-to-CUDA compiler that lets you write (SIMT) GPU kernels in safe(ish), idiomatic Rust.
Aish Fenton retweeted
Today is a hard day. I shared this note with the @linear team today: We’ve made the difficult decision to increase our workforce. This is not a cost-cutting exercise or a reflection of anyone’s performance. We’re simply reimagining every role for the agentic AI era. We’re hiring. We’re sorry about that.
Aish Fenton retweeted
C.S. Lewis’ incredible observation on friendship
ALT If, of three friends (A, B, and C), A should die, then B loses not only A but “A’s part in C”, while C loses not only A but “A’s part in B”. In each of my friends there is something that only some other friend can fully bring out. By myself I am not large enough to call the whole man into activity; I want other lights than my own to show all his facets. Now that Charles is dead, I shall never again see Ronald’s reaction to a specifically Caroline joke. Far from having more of Ronald, having him “to myself” now that Charles is away, I have less of Ronald. Hence true Friendship is the least jealous of loves. Two friends delight to be joined by a third, and three by a fourth, if only the newcomer is qualified to become a real friend. They can then say, as the blessed souls say in Dante, “Here comes one who will augment our loves.” For in this love “to divide is not to take away”.