Check out our Assistant Professor position @ucsd_cse!
https://t.co/rqaQVdwApW
Since it is 2026, it is worth quoting explicitly from the ad that this not AI-specific: “The department is looking for exceptional candidates in all areas of Computer Science and Engineering.”
We built LemmaScript with the *help* of LLMs, but it was carefully guided design-wise from day 1 by an actual PL expert. Every decision was scrutinized, every TypeScript feature we support was reasoned about. We threw away so many sessions, so many implementations. We ran hundreds of case studies (dozens of them are open) as we expanded the supported subset.
Then, we took a different approach to LemmaSwift. We wanted to see if, given what we did with LemmaScript, an LLM could create a useful and coherent formal verification toolchain for Swift.
When I tell you LLMs are NOT good at novelty yet. My god. We're on version 3 and this thing isn't ready to go out yet.
Starting over now with the same principled approach we used for LemmaScript. Literally starting from first principles, with hawk-eyed guidance, scrutiny on every decision, WE will be guiding the subset, and only using LLMs to speed up small chunks of implementation that WE decide on.
The models are SO good. And yet they are SO bad.
The surreal experience of trying to debug a robot for a personal project and getting frustrated and just plugging in a camera and telling Codex to figure it out
Some of you have been asking very politely. Some of you have just been asking.
GPT-6.2 Astra and Sol are now available.
Super efficient and built to perform where it counts—the real world.
https://t.co/qr2iP0GpfX
Who else is working on making AI better at cybersecurity and thinks today's evals aren't enough? I'd love to chat and compare notes!
I'm especially interested in brainstorming with people who are:
- Training AI models, e.g., frontier labs and large internal teams looking to improve cyber capabilities or model safety
- Building, deploying, or securing AI agents, e.g., building an AI SOC in-house or deploying a vendor's, hardening a widely used agent harness, or evaluating security agents. For background, our team publishes BotsBench (cyber investigation evals for AI models and harnesses) and builds @louie_ai (the first AI investigation harness to publicly beat and then clear Splunk's Boss of the SOC).
We're preparing new tooling and community resource launches around cyber evals, and investing even more here in 2027. Much of our work is public, but we also work privately with teams tackling these problems or looking for solutions.
If this sounds like you, or someone you know, I would to love to chat! Please reshare to help us reach the right people. And if you'd be open to a short conversation, please comment, DM, or even just "like" this post.
Thanks all!
I love exercising every day.
My secret is to limit myself to 20 minutes a day, and if I break the streak, I just add those 20 minutes to the next day.
I'm happy to share that today, I will be exercising for three weeks
I feel like every day I need to update my priors on what's hard, whats possible, and what might be worth my time.
Meanwhile, 99% of the world is operating on 2019 truths.
And most of the remaining 1% are programmers, with their own limited scope of what is valuable in reality.
Which leaves unprecedented upside for the rest of us.
LFG.
In order to kill Microsoft Word for lawyers, you must do the following:
- You can move one picture without the whole document rearranging itself.
- The cross-reference doesn't say "Error! Reference source not found." in the copy that went to the client.
- Pasting from an email doesn't drag in Calibri 11, three new styles, and a gray highlight nobody can remove.
- Two associates can edit the same document without one of them getting a file called "SPA_v14_FINAL_final_JM_edits(2)_USE THIS ONE.docx."
- When a deal blows up, the metadata doesn't show opposing counsel the comment "lol they'll never catch this."
- The signature block stays on the same page as the last sentence of the agreement.
- Footnote 14 doesn't vanish when the partner deletes a paragraph three pages earlier.
- "Accept all changes" doesn't also accept the one change everyone agreed was a terrible idea.
- Bullets don't turn into a different, slightly larger bullet when you hit enter.
- The table of contents doesn't list "Section 4.2(a)(iii)" as page 1 and every other section as page 1 too.
- Exhibit A is still Exhibit A after someone adds a new Exhibit A.
- Hitting "Restrict Editing" means something, unlike the last 20 years.
- Word doesn't crash at 2am, then recover a version from 4pm with "unsaved changes."
- Track changes doesn't record the partner reformatting all 80 pages as 6,000 changes.
- Track changes survive 6 rounds, 3 firms, and one client who "just made a few edits" with tracking off.
- Auto-numbering never restarts at 1 because someone pasted in a clause from a 2011 precedent.
- Comparing two drafts doesn't flag 400 changes when someone moved a space.
- Every defined term gets tracked. If "Company" is used before it's defined, you yell at me before opposing counsel does.
- The table of authorities builds itself, and it's right.
- It opens on a 2017 Dell, through Citrix, on hotel wifi, at 11:58pm, while the partner is "just checking in."
- Redlines, just do redlines well.
@mattpiccolella, the legal profession has waited 40 years for a savior. It's you now. Godspeed.
Coding agents made passes and proofs cheap. The hard part now is verifying the model and assumptions the proofs rely on, and making sure they match reality.
Full write-up: https://t.co/CEHU9qkIKh
Can a formally verified C compiler be as fast as gcc?
CompCert’s optimizations come with a machine-checked proof that they don’t change your program’s semantics. The catch? The compiled code runs ~28% slower than gcc -O2.
We tried closing that gap with AI agents 🧵