sandkoan

Yeah, users can set their own list of redacted words.

Good idea! Adding that as we speak.

You're absolutely right. The comedic relief scene is almost necessary to avoid us being stretched too taut after the previous. Thomas de Quincey has a great piece reflecting on the very same: http://www.shakespeare-online.com/plays/macbeth/knockingatga...

Something which may not be known to many, McCarthy was also quite interested in math, philosophy, cognition. A piece of his non-fiction work which made a lasting impression on me was "The Kekule Problem" [0], exploring the power of the unconscious, combative with consciousness, as an unveiler of Truth.

0: https://nautil.us/the-kekul-problem-236574/

Presumably, they're referring to chunking up the data into discrete semantic units—smaller vectorizable subsections (e.g., paragraphs) more precisely capturing different parts of the data.

Isn't it? Wrangling a poorly understood system with partial knowledge and experimental skill seems like engineering to me.

The imagery you just evoked reminds me of Bradbury's short story, "The Veldt." [0]

[0]: https://repositorio.ufsc.br/bitstream/handle/123456789/16372...

I recall Ben Horowitz saying a16z was a conscious nod to i18n in some interviews.

Bubeck has clarified that the "1 trillion" number he was throwing around was just a hypothetical metaphorical—it was in no way shape or form implying that GPT-4 has 1 trillion parameters [0].

[0] https://twitter.com/SebastienBubeck/status/16441515797238251...

I'm curious as to what people do currently, without runhouse. And how does runhouse work, under the hood? Could I trivially configure model serving without resorting to FastAPI or some such?

When you say you wish they had a "better story for local development," what do you mean? What do you wish for?

You may want to look at run.house [0] for a pretty powerful solution to many of these problems.

[0] https://github.com/run-house/runhouse

I only brought up the Gladwell post to clarify the scale at which these institutions operate. There is no question that they have the money, and, if they chose, could sustainably provide scholarships to some subset of the particularly meritorious. Nothing stops them from soliciting donations for a new merit fund just as they do for their chairs and libraries.

I'm not saying everyone deserves a merit scholarship—but if the point of university is to nurture the best and brightest, then the best and brightest shouldn't have to pay such exorbitant fees. Afterall, they're the reason the colleges maintain their sterling academic reputations.

If you have a household income of 250k—of which your take home is ~160k—and Columbia subtracts another 90k, that leaves you with 70k to pay for everything else. In expensive locales, that's not nearly enough to pay for the education of two children without mortgaging your house. If said best and brightest are unable to attend due to such circumstances, a merit scholarship seems in order—otherwise, you, as a college, are admitting that money means more to you than education, and that you're willing to take someone who isn't nearly as good (there's a reason you didn't admit them initially) simply because they're capable of paying. Not everyone who is admitted can afford to attend at the price of admission demanded by these colleges. Again, I speak from experience.

No, I'm criticizing their decision to only offer need-based scholarships when they could just as well provide both—if they really cared about academic strength, then they could put their money where their mouth is, and provide meaningful merit-based scholarships.

Princeton, for instance, is capable of letting in every student for free, and still not dipping into their endowment [0]. Why don't they do so? Because it would lead to lower tuitions across the board, and that's intolerable.

> if you're admitted, then you're going to get whatever amount of scholarship you need to go.

No, you get whatever amount of scholarship they think you need to go. Their understanding of your financial situation will be unsurprisingly self-serving.

[0] https://malcolmgladwell.bulletin.com/princeton-university-is...

The Ivies and adjacent elite colleges are more than capable of giving out academic scholarships—they simply refuse to because that would lead to a race to the bottom in terms of how much money they'd be able to extract from top talent (who could easily negotiate colleges against each other for lower and lower tuition). And, from experience, Colubmia's defection hardly counts for much (which is why, I suspect, it's permitted): even their "prestigious" Rabi Scholarship yields no reduction in a mind-boggling $90,000 tuition—just a minute research stipend. They want to make you feel special while ensuring you still pay full price.

I'm actually building this very thing—shoot me an email at govind <dot> gnanakumar <at> outlook <dot> com if you'd like to be a beta tester.

I'm actually building this very thing—shoot me an email at govind <dot> gnanakumar <at> outlook <dot> com if you'd like to be a beta tester.

Presumably, because they trained Copilot on billions of lines of, often licensed, code (without permission), that Copilot has a tendency to regurgitate verbatim, without said license.

Might this be what you're looking for: https://github.com/bigscience-workshop/petals ?

I'd honestly argue that he might not have even needed OpenAI embeddings—any off-the-shelf Huggingface model would've sufficed.

Because of attention mechanisms, we no longer so heavily depend on the existence of those "pointy words," so generally, Transformers-based semantic search works quite well.

See also Popper's paradox of tolerance: https://en.wikipedia.org/wiki/Paradox_of_tolerance

Huh, never really noticed that on my end. I'm curious as to what question you asked....

If the rewind feature set—especially the search through screen recording aspect—is the most compelling, then perhaps something I've been working on may be of use. It can basically do everything rewind can (though audio support is a WIP), doesn't cost anything, and feedback would be really helpful.

Shoot me an email at govind <dot> gnanakumar <dot> com if you're curious or interested in beta testing it.

I've been working on a semantic search engine for code that's particularly useful to those navigating unfamiliar codebases. Just write an English description of whatever you're looking for, and it'll show you the most relevant/similar functions from the repository.

Afterall, it's often useful to see how things are done, and why. Search at least makes possible the former.

Email me at govind <dot> gnanakumar <at> outlook <dot> com if you'd like to try out a beta—on account of potentially being hugged to death, I don't want to publicly display the link.

No, the basic model already exists. You could throw together something like this yourself: encode your documents with a Transformers model into a series of vectors. Then you're merely a nearest neighbor search away from finding the most semantically relevant documents. Feed those documents to GPT-3 or some such LLM as contextual state, along with the query, et voila! you have Q&A on documentation.

For a distributed computing/BitTorrent-style method of running these LLMs, see: https://github.com/bigscience-workshop/petals.

I've actually been building something like this: a desktop app that basically records everything you see and hear, and makes it searchable. Maybe we could team up: my email is govind <dot> gnanakumar <at> outlook <dot> com.

This might be what they're referring to: https://regex101.com/.

I, too, am an aphant (which I believe affects only 3 to 10 percent of the population), but I understood this to still apply to us: the picture is merely replaced with the small kernel of a concept it represents.