Fraser (@FraserLeeee) profile picture

GPU Architectural Performance team @ Apple BSc compsci+physics @ McGill, '25

Boston, MA
Joined July 2018
Photo by Fraser (@FraserLeeee)Photo by Fraser (@FraserLeeee)Photo by Fraser (@FraserLeeee)Photo by Fraser (@FraserLeeee)Photo by Fraser (@FraserLeeee)Photo by Fraser (@FraserLeeee)Photo by Fraser (@FraserLeeee)Photo by Fraser (@FraserLeeee)Photo by Fraser (@FraserLeeee)Photo by Fraser (@FraserLeeee)
Turns out if you keep piling on more neutrons it eventually becomes stable again
116
194
39
3,660
123,635
Islands of stability are created over millions of years as nucleon counts drift over a stationary mantle plume.
2
1
250
11,920
When laity ask frontier models for medical advice, they'll hedge ("I am not a doctor", etc.). When professionals ask the same questions in different terms, the models answer confidently. When hedging, how aware are they of deceptively underrepresenting their own capability?
189
Mildly surprising that I haven't yet received any emails trying to prompt-engineer me.
1
174
Congratulations to OpenAI, Astra seems incredible. To me, this paragraph reads like an agent regurgitated part of its prompt into the final output. > The aim is to collect and explain the mathematical thinking behind the proofs in a readable, high-level narrative. Rather than simply reproducing their finished presentations, the chapters emphasize the connections, intermediate discoveries, and sustained detours that clarify why the successful arguments work. Current models do this sort of thing a lot. So do students, trying to cheese points on an essay by badly guessing towards the teacher's password. I'm not happy that as more math problems crumble, the leading lab still doesn't have the control needed to curtail such an obvious First Order Optimal Strategy. (This is an uncharitable reading, it's possible this paragraph could end up written in such a weird way for benign reasons).
An internal version of our next major model produced 10 new results on long-standing open problems in mathematics and theoretical computer science, using roughly $2,000 worth of tokens at GPT-5.6 Sol API rates.
1
193
Claude, that's not a PCB. It's swiss cheese.
1
85
(Trying to route a board w/ Z3)
54
me: surprised by how much a few analog circuit components could cost digikey: yes, ofc, Sixteen Hundred C batteries
1
97
There was never a world where the general public maintained access to frontier models after frontier models became actually useful. People complaining about the Fable router, they should know they're just telling labs to quietly hobble models without any explicit notification.
1
72
Credit to Anthropic for knowing *exactly* what the reaction would be yet still a) disclosing fallbacks on the API b) displaying it loudly on their front-ends and c) including Fable in the cheap pleb plans in the first place (guaranteed negative attention while burning money)
72
Mourn the loss of readmefirst.txt, START_HERE.txt, and the total conventionless zoo now paved over by README.md
69
Oh yeah, no way 4.8 could have possibly inferred what I meant by “south” in this context 🙄
62
Hah. Not a trained bespoke thought-translator, just a little lIm told to do its best.
1
66
Anthropic's classifier is convinced the only reason anyone would write C and make system calls is for evil.
1
65
I literally cannot call fork, or ptrace, or kill children, or talk about leaked file descriptors.
1
85
This was fine as of ~1 week ago, fwiw
32
:let i=1 | %s/^\s*\zs- /\=printf('%d. ', [i, execute("let i += 1")[0])/g
66
Every ambulance should have a small fleet of reconnaissance drones
58
Bash + all unix utils is too powerful to audit or trust. Agent frameworks should replace it with something syntactically similar, but structured, sampled from the model such that python-in-heredocs, shell in strings, etc, are impossible.
1
63
Or just replace RL incentive to minimize number of tool-calls, with incentives to minimize number of groups of tool-calls, plus incentivize making every individual call atomic and interpretable.
1
45
You wrote a beautiful permission-scoped rewindable accountable Edit() tool. Your model prefers Bash(cat<<EOF > /tmp/foo.c\nbar\nEOF)
45