cjbprime

25 ms·

361.
by cjbprime 3y ago
It's a little complicated, because I don't know that Google has disclosed whether they have a multimodal LLM, or OCR software and other more standard machine learning hooked up to their text LLM. If it's true multimodal, I&#x
362.
by cjbprime 3y ago
And make sure never to re-use one lambda function or container to process more than a single customer's data, to avoid the RCE gaining access to whatever file next gets uploaded :)
363.
by cjbprime 3y ago
Lots of retrospectively-good predictions in there, but this was the one I undervalued the extent of: > We also have a bad habit of changing the definition of machine intelligence when a program gets really good to claim that the problem
364.
by cjbprime 3y ago
I think you found the "good reason I am not thinking of", nice work.
365.
by cjbprime 3y ago
Had a thought, maybe someone here will know the answer -- has anyone trained a voice system on how long to wait before replying to an utterance? It occured to me that if your STT hears "hey how's it going" it should reply imm
366.
by cjbprime 3y ago
Somewhat covered by other comments, but what I would want to see to evaluate a project like this is your current response latency and your goal latency for each step, or ideas about reducing it.
367.
by cjbprime 3y ago
> There are two possible explanations for this effect: > 1. The tech industry has accidentally invented the initial stages a completely new kind of mind, based on completely unknown principles, using completely unknown processes that
368.
by cjbprime 3y ago
> You're also better off with a newer consumer-level processor, Strong disagree on this one. You can't even put 64GB of DDR5 into the latest consumer-level processors without causing bandwidth to halve down to DDR4 speeds (bec
369.
by cjbprime 3y ago
I've been thinking about buying one of these, they show up on eBay as CPU+mobo+RAM combos as used cloud provider machines. Each CPU is 64C/128T, instead of the 15C exposed in the post's cloud VM. The rackmount cloud machine
370.
by cjbprime 3y ago
I didn't see any mention of AI tech -- using embeddings to automatically surface relevant documents, a chat mode to answer questions about them, etc. It seems like the legal industry always has a great deal of inertia around tools, suc
371.
by cjbprime 3y ago
I think I read that AmpereOne isn't Neoverse.
372.
by cjbprime 3y ago
This means that either you don't get to see what your document is going to look like while you write it, or you have two copies of the document open at the same time (one showing the preview) and move your eyes between them, and neithe
373.
by cjbprime 3y ago
Well, used to be a Very Hard Problem. Now you can just test it: train a model on the textbook and ask it to solve many problem challenges that it hasn't seen before. After you get a score for the model trained on the whole textbook,
374.
by cjbprime 3y ago
Most of the time you want to confirm that you're talking to someone from a given context -- they own a specific Twitter account, or you met them at a party last week, or they sent you an email or were present in a meeting that you want
375.
by cjbprime 3y ago
Try different phrasing, like "Did your answer follow all of the criteria?".
376.
by cjbprime 3y ago
I'd like to see new papers link to a fuzzbench.com experiment showing that they are the new rank 1 fuzzer (the experiment would be conducted by the Fuzzbench infrastructure), if they're going to claim it in a paper. Feels like eve
377.
by cjbprime 3y ago
I think that some LLMs (mainly just GPT-4) should be considered as refutations to the Stochastic Parrot idea, which was published in March 2021 and claims no LLM can have "any model of the world". It was a reasonable (though perh
378.
by cjbprime 3y ago
> you want a document, you are two steps away from it in an easy to find way This isn't a response to the parent commenter's point, right? They were describing how many projects have items where a resource easily fits within t
379.
by cjbprime 3y ago
You should try web searching for your question before asking it. If it shows up, don't ask it. If it doesn't show up, and the complaint is that it's possible to solve the question by studying past answers to different questio
380.
by cjbprime 3y ago
I can be sympathetic to many folks who have publicly dismissed LLMs as vapid hype in past months -- tech has claimed many fundamental breakthroughs before, it was reasonable to expect this would be another illusory one too -- but come on, e
381.
by cjbprime 3y ago
The bandwidth is fastest with one single-rank DIMM per channel, and there are four slots and two channels, so you ideally want two single-rank (chips on one side) DIMMs total. It'll get slower with 2DPC (DIMMs Per Channel), and slower
382.
by cjbprime 3y ago
Reasons to still prefer an M2 for a HEDT situation, despite the article claims: * The latest Intel/AMD desktops only have two channels of DDR5, and if you put as little as 64GB RAM in them they can drop to DDR4 speeds (2DPC, dual-rank)
383.
by cjbprime 3y ago
It felt like strong "I am recently divorced, and also a parent of young children" vibes throughout. There was the value comparison to a large TV and surround sound system, which is only valid if both your home theater and AR devic
384.
by cjbprime 3y ago
You don't seem like someone who had used GPT3, let alone GPT4. It has an excellent command of language. I doubt there's a single question requiring "human linguistic ability" you can come up with that GPT4 will do worse
385.
by cjbprime 3y ago
The relevant performance metric here is memory amount and bandwidth.
386.
by cjbprime 3y ago
I agree, though that might be where the survivorship bias comes in -- it seems like Elm's ideas have been massively successful and I would say are now the most common way to produce webapps. But people choose to take the ideas while di
387.
by cjbprime 3y ago
Elm seems like the best candidate?
388.
by cjbprime 3y ago
And an alternative to push marketing -- rate-limit the site to a capacity you can afford, with a prominent message saying that you enjoy working on the site and would like to keep it going but can't afford to, with something like a Ven
389.
by cjbprime 3y ago
> sloppy conference planning FWIW, the quoted thread makes it sound like the actual conference planners were no more than messengers here, so it may be fairer to summarize as ".. over capricious meddling in conference planning"
390.
by cjbprime 3y ago
https://github.com/MineDojo/Voyager/blob/main/voyager/prompt...

More ›