alansaber

7 ms·

121.
by alansaber 1mo ago
The overarching geopolitics have always taken precedent over the preferences of the mathematical community as far as I can tell
122.
by alansaber 1mo ago
For sure. Even if it wasn't a measure to avoid copyright, you pre-process LLM training data to remove errors, characters that can't be tokenized, etc etc. Doing so with another LLM has been standard for a while.
123.
by alansaber 1mo ago
This is just a blogpost, rather than a major announcement (FLT proof was closer to the latter than the former)
124.
by alansaber 1mo ago
Yes I agree. AI has fundamental problems but a lot of "AI sucks" is the huge, trivial capability for misuse.
125.
by alansaber 1mo ago
Hot take but I don't mind startup founders being plucky and a bit arrogant/naieve. They need it.
126.
by alansaber 1mo ago
AI copy from the AI company incubator is hardly the worse sin
127.
by alansaber 1mo ago
No but it reinforces the value of the YC brand. VC is always looking for an edge
128.
by alansaber 1mo ago
Usually the rule of a CTO/CIO is to foster innovation. Getting a custom solution from a big name VC helps then fulfil their JD within the company.
129.
by alansaber 1mo ago
There are far worse gigs than VC backed founder.
130.
by alansaber 1mo ago
There are good arguments for subagents/multi-agents. But the added overhead is usually massive.
131.
by alansaber 2mo ago
For our data centres we must go to the one place not yet corrupted by capitalism- space!
132.
by alansaber 2mo ago
It's dubiously collected data the whole way down. This is the AI industry.
133.
by alansaber 2mo ago
The question is not, will mistral be able to create the best models (seems bloody unlikely). The question is, will mistral have enough expertise to dominate its business use-case, which is the model infra/deployment market within the E
134.
by alansaber 2mo ago
To be fair, Switzerland is really expensive.
135.
by alansaber 2mo ago
"something that will work reliably and predictably for years" is not really the class of product being sold, unless you're using a fine-tuned SLM to do something like classification. The vast majority of work being done with
136.
by alansaber 2mo ago
Stay out of the race sounds insane when AI is, and will be, a common commodity. China is getting great bang for their buck on AI training, why not the same for Europe?
137.
by alansaber 2mo ago
Holy oversimplification. Do you get all your information from the internet?
138.
by alansaber 2mo ago
They've never made a near-frontier model. Which is a valid criticism for a model training company
139.
by alansaber 2mo ago
There is nothing exciting about mistral but they're the only european ai lab i've even heard off (jeppa doesn't count)
140.
Optimising Harness Self-Recursion(lexifina.com)
1 points by alansaber 2mo ago|0 comments
141.
by alansaber 2mo ago
Certainly agreed, but this is more about refining unconstrained vibe-lawyering (and the resulting misfilings and hallucinations) than automating judgements. Even if we had a perfect steamroller of strong types, there would still be a lot of
142.
The auto-formalization of mathematics versus law(lexifina.com)
3 points by alansaber 2mo ago|2 comments
143.
by alansaber 2mo ago
It's far from trivial to set up anything more formalised than a loose taxonomy. People have been trying for a few decades now. AI will help automate this formalisation with graphs etc but it's not good enough yet for you to give a
144.
by alansaber 2mo ago
Just as many junior devs are serving slop PRs. Unfortunately, you can only take a horse to water.
145.
by alansaber 2mo ago
We've rediscovered the need for academia from first principles
146.
by alansaber 2mo ago
Unfortunately framing everything as incentives is the only non-emotional way to analyse these meta-behaviours (get it?).
147.
by alansaber 2mo ago
No, that will be the gambling endemic in every aspect of our society.
148.
by alansaber 2mo ago
Actually, they all seem solely interested in the future. The bunkers are a drop in the bucket for them, so why not?
149.
by alansaber 2mo ago
Usually people only bother reading it if the outcome is interesting
150.
by alansaber 2mo ago
AI is good at volume, bad at accuracy and nuance. Of all the tasks to go through with a fine-toothed comb (ie, manually) legal is one.

More ›