
🇨🇦/acc, technoalchemist @chaosgremlinlab
Lethbridge, Alberta
Joined August 2008
- Tweets2.3K
- Following2.5K
- Followers438
- Likes2K
Chris Rabl retweeted
instead of giving your agents / llm whatever a soul.md, tell it to never use first person, that it is an llm, not an "i"
its amazing, never going back
is this how mathematicians felt last week
BREAKING: OpenAI says it found a critical flaw in Brotli, a widely used compression library built into software like Node.js, Bun and Envoy.
The bug could let an attacker take over a system just by sending it malicious compressed data, with no login needed. That works both ways: a hacker could hijack an unpatched server, or a malicious server could hijack an unpatched user’s device.
Chris Rabl retweeted
Craziest news of the week: Andon Labs has a product
the CrossX app built by @jtaverasv is phenomenal and "just works". if you have an Xteink X3 or X4 it's definitely worth a download. i'm super impressed by the OSS community around this hardware!
incoming @modal appreciation post: i freaking love this product. ridiculously easy to run experiments with serverless GPUs!
1
10
1,099
grok 4.5 is decent (not on the same level as opus) but limited by 500k context, where it wins is SPEED omg this thing is fast
1
1
304
Chris Rabl retweeted
I might be one of the few people who is most bearish on human research taste and bullish on automated research:
- "AIs can only do hyperparameter search" is mainly a skill issue with bad automated research setups.
- human taste is overrated, e.g. frontier labs / neolabs are doing pretty simlar things.
- human taste might win in a low-compute world, but not a high-compute world we're entering.
opus 4.7 is extraordinarily good at following precise instructions, unfortunately it is also extraordinarily good at following instructions precisely
112
thank you @modal for existing! serverless H200 instances with snapshots are so nice for quick experiments
1
67
been data mining my claude code transcripts to figure out why conversations "feel worse" lately. pulled out 264 "tilt" incidents across 127 sessions, ~9k tool calls and this is what i found...
1
1
121
in contrast to others i don't think the fix is more thinking tokens. tool selection and "thrash" detection are two places to start: make the model verify against actual source material before it starts claiming things, and detect when the model starts swirling and second-guessing
1
63









