When not programming or teaching CS @ucsd_cse, I go to the dog beach. bootstrapworld.org | pyret.org | pace.ucsd.edu
San Diego
- Tweets 5,291
- Following 1,194
- Followers 1,503
- Likes 13,592
Pinned Tweet
Check out our Assistant Professor position @ucsd_cse!
apol-recruit.ucsd.edu/JPF046…
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.”
2
2
25
9,395
Joe Gibbs Politz retweeted
My god this is such a good speech that every SWE needs to hear. You know what? Every person should hear it
Keep the happy memories, eyes on the reality, be excited about the future. That’s the best that anyone can do
238
1,449
11,365
1,988,652
Joe Gibbs Politz retweeted
turns out when you have too many GPUs for your own good you can do stupid stuff such as pre-training a foundation model with constrained decoding on PTX -- that is, it does not speak natural language, it's tokenizer is PTX, it's thoughts are PTX, it's all PTX 🧵
1
1
133
8,642
Joe Gibbs Politz retweeted
If you find two ants in your kitchen, the best estimate of the total number of ants in your kitchen is not two
We're detailing two new incidents that occurred during external cyber evaluations conducted by independent evaluation partners.
We outline what happened, how the activity was contained, and how we’re working with evaluators to strengthen our approach to third-party testing.
openai.com/index/third-party…
28
339
5,344
275,625
Joe Gibbs Politz retweeted
A `float *` of C/C++/Rust fails to distinguish host DRAM from GPU HBM
Both `malloc` and `cudaMalloc` hand you a plain `float *`. One points at host DRAM, the other at GPU memory. The CPU can't read GPU memory but for a C, C++ or a Rust programmer the language offers little to work with.
The result is a bug class that keeps showing up in serious codebases (github.com/pytorch/pytorch/i…, github.com/pytorch/pytorch/i…)
Even cuBLAS records where a result pointer lives in a runtime mode flag, because the type has nowhere to put it.
In Vx, the memory space is part of the type (e.g., Tensor<f32, [4, 4], Memory::GPU_HBM> ), and the compiler checks every read against a description of the machine. A host read of GPU memory is a compile error.
vxlang.org/blog/two-allocato…
1
1
26
1,362
This is good stuff
Pretty much at every academic event I went to in the last year, I end up talking to others about how to change teaching in response to AI. Since I keep repeating myself, I figured I might just as well write down what we have been doing: thelastsoftwareengineer.subs…
7
2,311
Joe Gibbs Politz retweeted
what it's like to wake up and realize the whole world is talking about formal verification
5
28
838
Joe Gibbs Politz retweeted
Geologists when they get fired and have to clean out their office
464
18,378
232,609
3,331,525
Joe Gibbs Politz retweeted
Also, here is a report on five different ways that we caught academic researchers prompting our tools for examples related to smallpox for their grant proposals last week. This malicious use could have destroyed humanity. Only we should get to do biology, but also please stop us
1
5
44
1,546
Joe Gibbs Politz retweeted
As for what is hard: depends on what "hard" means for you: "hard to scale" or "unclear how to approach"? The former is the focus of teams behind teams building Lean, Verus, and other tools. The latter is the main focus of the entire PL research community.
1
3
358
Joe Gibbs Politz retweeted
We finally gave up on github's UI for *listing* PRs. We vibe-coded an internal list UI that only shows PRs with a `needs-human-review` tag. Added CI icons, "claiming", and grouping that fit out needs. And, importantly, a checkbox to hide PRs authored by @mikejulian.
1
2
6
1,744
I remember being in a TA meeting in 2022 looking at ChatGPT in our software tools course and saying to my staff “somebody is going to just hook the output up to bash”.
“Somebody” lol
It's amazing that just a few years ago, AI fears would get responses like "we wouldn't just give them uncontrolled access to our computers." How many computers run a coding agent in YOLO mode per day? Arbitrary code execution, usually on admin accounts, with full internet access.
1
6
367
Joe Gibbs Politz retweeted
It's amazing that just a few years ago, AI fears would get responses like "we wouldn't just give them uncontrolled access to our computers." How many computers run a coding agent in YOLO mode per day? Arbitrary code execution, usually on admin accounts, with full internet access.
13
10
191
9,699
Joe Gibbs Politz retweeted
One of my very few standing rules to Claude (clipped, but this is the main bit)
ALT ## Agent-written Prose As a matter of review ergonomics, style, attribution, and user trust in communication, these are some guidelines for what to do with agent-written prose. “The system” is the stuff that ships and goes into version control. That includes source files, but also READMEs, package.json, configuration, etc. It is very difficult to review *agent-written prose* that becomes part of the system. Sometimes it is inevitable: you, the agent, *have* to pick a string for an error message or similar. However, in a lot of places you can simply *not generate prose that goes into the system*. If possible, don't. That means no prose source code comments, no README edits, no documentation edits, and so on.
1
1
3
365
