theahura

12 ms·

271.
by theahura 1y ago
> Topology is useful especially when you cannot trust lengths, distances angles and arbitrary deformations But...you can't. The input data lives on a manifold that you cannot 'trust'. It doesn't mean anything apriori
272.
by theahura 1y ago
> Topological transformation of the manifold happens during training too. That makes me wonder: how does the topology evolve during training? If you've ever played with GANs or VAEs, you can actually answer this question! And the an
273.
by theahura 1y ago
This is way too rigorous. You can absolutely have theories that lead to useful results even if they aren't as predictive as you describe. Theory of evolution for an obvious counterpoint.
274.
by theahura 1y ago
Chris Olah? One of the founders of Anthropic and the head of their interpretability team?
275.
by theahura 1y ago
Transformers learn embedding representations of tokens, which are easily mapped into a space. Similar tokens are mapped to similar places on the space. The fully connected layer at the end of each transformer block defines a transformation
276.
by theahura 1y ago
I say this as someone who has been in deep learning for over a decade now: this is pretty wrong, both on the merits (data obviously lives on a manifold) and on its applications to deep learning (cf chris olah's blog as an example from
277.
Deep Learning Is Applied Topology(theahura.substack.com)
508 points by theahura 1y ago|183 comments
278.
xAI screws up it's system prompt, again(theahura.substack.com)
3 points by theahura 1y ago|0 comments
279.
Paper Review: Neural Turing Machines(theahura.substack.com)
1 points by theahura 1y ago|0 comments
280.
by theahura 1y ago
Generally I like precommitment, even without the government enforcement. It keeps people honest. For most people, the shame of being called out is enough to have the precommitment have some effect (For OpenAI, with billions of dollars on
281.
OpenAI struggles to escape its non-profit chains(theahura.substack.com)
4 points by theahura 1y ago|2 comments
282.
The Board Was Right to Fire Sam Altman (2023)(theahura.substack.com)
2 points by theahura 1y ago|0 comments
283.
by theahura 1y ago
Quoting from the article: > In the immediate short term, I would love to see an ambitious startup take a crack at the credentialing problem using exams. I think it's extremely doable to make some exams; the hard part is the same chi
284.
Colleges Have an Education Problem(theahura.substack.com)
1 points by theahura 1y ago|3 comments
285.
by theahura 1y ago
who hurt you?
286.
by theahura 1y ago
Open to suggestions for an edit or will let dang do it (sorry for creating more work for you @dang). That said, I do stand by the title. This EO demands the use of military to 'prevent crime' and 'assist in law enforcement&#x
287.
by theahura 1y ago
From the order: > Sec. 4. Using National Security Assets for Law and Order. (a) Within 90 days of the date of this order, the Attorney General and the Secretary of Defense, in consultation with the Secretary of Homeland Security and t
288.
The latest executive order creates a police state(whitehouse.gov)
61 points by theahura 1y ago|20 comments
289.
Congress moves forward with new attempts at internet censorship(techdirt.com)
20 points by theahura 1y ago|0 comments
290.
by theahura 1y ago
I'm doubtful. I don't see any comma.ai robotaxis coming online any time soon. When you say "doing fine", what do you mean?
291.
by theahura 1y ago
O did it get flagged? I was wondering! I saw this was doing numbers through substack while I was at dinner and then it suddenly seemed to stop EDIT: lol looks like it got flagged again
292.
I wouldn't get into a Tesla Robotaxi(substack.com)
48 points by theahura 1y ago|32 comments
293.
Towards a theory of model complexity in deep neural networks(theahura.substack.com)
2 points by theahura 1y ago|0 comments
294.
What is Tesla's strategy now?(theahura.substack.com)
5 points by theahura 1y ago|0 comments
295.
by theahura 2y ago
I mention distribution in the post, and reasons to be skeptical of that as the primary driver (though I agree that may be the case) Re talent, I'm not sure how big windsurf is, but aren't these teams generally quite small? $3b for
296.
by theahura 2y ago
(I mention in the footnotes that this is likely a stock deal!)
297.
by theahura 2y ago
Thanks for sharing, super interesting!
298.
by theahura 2y ago
(I most recently used cursor in October before switching to Avante, so I suspect I've experienced the version of the tool you're talking about. I mostly didn't use the autocomplete, I mostly used the chat-q&a sidebar.)
299.
by theahura 2y ago
Strong plus one. This is more or less what my company is working on -- more and more ostensibly nontechnical people are able to contribute to codebases with seasoned engineers.
300.
by theahura 2y ago
Ok thanks, that was my follow-up -- I assumed that airgap implementations are significantly worse because they can't back into Claude or Gemini

More ›