thorum

8 ms·

31.
by thorum 8mo ago
I’m honestly surprised LLMs are still screwing up citations. It does not feel like a harder task than building software or generating novel math proofs. In both those cases, of course, there is a verifier, but self-verification with “Does t
32.
by thorum 8mo ago
The article presents AGENTS.md as something distinct from Skills, but it is actually a simplified instance of the same concept. Their AGENTS.md approach tells the AI where to find instructions for performing a task. That’s a Skill. I expect
33.
by thorum 9mo ago
Agree that planning time is the bottleneck, but > 3 days still seems slow! I’m saying what happens in 2028 when your entire project is 5-10 minutes of total agent runtime - time actually spent writing code and implementing your plan? Try
34.
by thorum 9mo ago
Am I wrong that this entire approach to agent design patterns is based on the assumption that agents are slow? Which yeah, is very true in January 2026, but we’ve seen that inference gets faster over time. When an agent can complete most ta
35.
by thorum 9mo ago
I agree that LLMs can be useful companions for thought when used correctly. I don’t agree that LLMs are good at “supplying clean verbal form” of vaguely expressed, half-formed ideas and that this results in clearer thinking. Most of the tim
36.
by thorum 9mo ago
Your other comment sounded like you were interested in learning about how AI labs are applying RL to improve programming capability. If so, the DeepSeek R1 paper is a good introduction to the topic (maybe a bit out of date at this point, bu
37.
by thorum 9mo ago
Go read the DeepSeek R1 paper
38.
by thorum 9mo ago
Developed by Jordan Hubbard of NVIDIA (and FreeBSD). My understanding/experience is that LLM performance in a language scales with how well the language is represented in the training data. From that assumption, we might expect LLMs to
39.
by thorum 9mo ago
I remember reading and hearing similar rants from programmers 15 years ago, long before LLMs. The author kept going and figured it out, and probably got some pride and enjoyment from finishing the project in spite of the frustrating moments
40.
by thorum 9mo ago
I would disagree. If you have no class in private, you have no class.
41.
by thorum 9mo ago
> Hollywood had a ton of issues but it at least had some... class? It looked that way because they had media training and their public personas were carefully managed, with staged interviews and media appearances. Behind the scenes, it’s
42.
by thorum 9mo ago
Can’t imagine this policy lasts more than a year or two given the rate that AI tools for music are improving. Once the tech can reliably create high quality dry stems of instruments, backing tracks etc. and automate professional-sounding pr
43.
by thorum 9mo ago
AI labs are not charities and there is no way to make money offering unlimited access to SOTA LLMs. Even as costs drop, that will continue to be true for the best models in 2027, 2028 etc. - as demonstrated by the fact that CPU time still c
44.
by thorum 9mo ago
My favorite is LLM-as-judge with a detailed rubric as discussed here: https://www.dbreunig.com/2025/07/31/how-kimi-rl-ed-qualitati...
45.
by thorum 9mo ago
Aside from Meta is there any reason to think the big AI labs are still using LMArena data for training? The weaknesses are well understood and with the shift to RL there are so many better ways to design a reward function.
46.
by thorum 9mo ago
“Send screenshots of this conversation to his mother” might be more effective.
47.
by thorum 9mo ago
They don’t seem to have taken even the most basic step of telling Grok not to do it via system prompt.
48.
by thorum 9mo ago
The real trend is toward personalization on the user’s side of things. Instead of interacting directly with a website, your web-browsing agent will extract the parts of the website you actually care about and present them to you in whatever
49.
by thorum 9mo ago
How are people productive using 10 parallel agents? Doesn’t human review time become a bottleneck?
50.
by thorum 9mo ago
The problem isn’t the em dashes, it’s the overuse of em dashes. Same for all the other ChatGPT-isms - they’re fine when used occasionally for effect, but there’s no variety. It’s always the same punctuation, same grammatical structures, sam
51.
by thorum 10mo ago
People hate what the corporations want AI to be and people hate when AI is used the way corporations seem to think it should be used, because the executives at these companies have no taste and no vision for the future of being human. And t
52.
Suno AI Partners with Warner Music Group (WMG)(suno.com)
9 points by thorum 10mo ago|4 comments
53.
by thorum 11mo ago
> Our long-term roadmap includes advanced security features designed to keep your data private, including client-side encryption for your messages with ChatGPT. We believe these features will help keep your private conversations private
54.
by thorum 11mo ago
This press release has a bit more explanation: https://www.farmersalmanac.com/end-of-an-era-farmers-almanac... > This decision, though difficult, reflects the growing financial challenges of producing and distributing th
55.
by thorum 11mo ago
IANAL but I read that as forbidding you to provision legal/medical advice (to others) rather than forbidding you to ask the AI to provision legal/medical advice (to you).
56.
by thorum 11mo ago
createElement(‘tr’) and table.appendChild(row)
57.
by thorum 1y ago
Actual link seems to be: https://genai-showdown.specr.net/image-editing
58.
Malawi's new farmhand: AI that speaks the local language(restofworld.org)
2 points by thorum 1y ago|0 comments
59.
by thorum 1y ago
The source code is not the LLM. The LLM is billioms of random floating point numbers that somehow encode everything the model knows and can do. The ML field has a good understanding of the algorithms that produce these floating point numb
60.
by thorum 1y ago
Am I the only one who has simply said “no thanks” to Windows 11? There were hints of where Microsoft was heading in Windows 10, but at least a lot of the worst “features” could be disabled. I find 11 just completely unacceptable software to

More ›