Kyle Maxwell

@fchollet Agree, but whats your over/under for hourly ongoing inference costs to get human-level knowledge work assuming we had such a system?

I believe that if you understood intelligence, you could build it in software form on a $1M budget, including training. It would not need to be trained on the entire Internet, nor on the thought process of thousands of experts.

@thdxr RSC subtweet

i think i've seen this enough times that i might consider it a pattern: let me start with an example - in our space this potentially interesting idea called "infra from code" popped up a few years ago the idea was letting you mix infra and application code together - could build things with an impressive lack of boilerplate we were definitely tempted in this direction because it solved some ugliness in sst and our stuff looked cumbersome by comparison ultimately we refrained and these IfC frameworks have seen little adoption since - i think this fits into a larger pattern in software people will observe a boundary. and they will observe all the negative things the boundary causes - extra boilerplate, duplication and even organizational problems they will build a solution to erase the boundary which immediately seems successful because any 0 to 1 example looks better but these types of solutions rarely seem to stick it turns out most boundaries exist for a good reason - they provide explicitness, blast zones for refactoring, and a way to decompose your systems to keep them manageable the downsides people observe are real - but deleting the boundary is actually the laziest solution. you can instead improve how easily people can work on either side of the boundary and how smooth the interface between them is think of things like hasura which tries to eliminate your api vs tRPC which keeps it but makes it smoother to consume from the client - the latter is way more successful so now i've become extra skeptical of boundary eliminating tech

@swyxio Closer to the former than the latter

Ok hear me out—downhill mountain biking biathalon.

@Scobleizer @GoDaddy If you’ve done your trademark homework, there might be an angle there. Can’t really imagine less than 28k in legal fees tho.

@sampullara Clbuttic!

openai code censorship! hilarious ham-fisted attempt to somehow stop something but it is unclear what:

@AdamRackis @mattpocockuk @EffectTS_ It's how you write Scala in JS/TS

@mattpocockuk @EffectTS_ What exactly does this Effect library _DO_? The docs are super unclear. Is this a runtime? Is it just a set of typed utilities to help you do gross async operations while gathering up the results?

Feels like you could start being productive in @EffectTS_ with only 10-15 functions, but I have no idea yet which of the 500 or so of them I should choose. Still, my investigations continue.

@swyxio need the devin hot take

@mattknox @sampullara What if twitter DMs had more than 1 database server

@sampullara we'd have had infra challenges, but I think it would have been a winner.

what if vine had let people make longer videos

@antirez Every Star Destroyer programmed by AI, every rebel ship coded by hand?

Literature saved me from believing many trivial things that IT people sometimes tend to believe, such as the idea that if programming is eventually solved by AI, it will no longer be worthwhile for humans to write code.

@KaroroMitchelle This took a turn ... for me. I thought you were saying that he got the same present his brother always got her, which might be the sweetest thing in the world.

for me. Am very angry and feel degraded how could he buy both of us the same present? Am l missing something here?

My partner's brother passed away late last year and the wife phoned yesterday asking me to thank my partner for buying her a valentine present. I pretended to have known about it and she went on to tell me what it was and guess what ladies it’s the same present he bought

@typedfemale No one expects the bitter lesson, even though that's the point

(also it should make you very skeptical of anyone who's pushing a fancy architecture for intra-document or intra-codebase long-context solution)

i get more excited about the near future from the gemini 1.5 long-context reasoning demos than from sora (of course both are incredibly impressive)

@JeffDean How does it do on large code bases not in the training set?

Analyzing and understanding complex code bases The model’s ability to quickly ingest large code bases and answer complex questions is pretty remarkable. three.js is a 3D Javascript library of about 100,000 lines of code, examples, documentation, etc. (threejs.org) With this codebase in context, the system can help the user understand the code, and can make modifications to complex demonstrations based on high-level human specifications (“show me some code to add a slider to control the speed of the animation. use that kind of GUI the other demos have”), or pinpoint the exact part of the codebase to change the height of some generated terrain in a different demo. Watch this video! youtube.com/watch?v=SSnsmq…

Gemini 1.5 Pro - A highly capable multimodal model with a 10M token context length Today we are releasing the first demonstrations of the capabilities of the Gemini 1.5 series, with the Gemini 1.5 Pro model. One of the key differentiators of this model is its incredibly long context capabilities, supporting millions of tokens of multimodal input. The multimodal capabilities of the model means you can interact in sophisticated ways with entire books, very long document collections, codebases of hundreds of thousands of lines across hundreds of files, full movies, entire podcast series, and more. Gemini 1.5 was built by an amazing team of people from @GoogleDeepMind, @GoogleResearch, and elsewhere at @Google. @OriolVinyals (my co-technical lead for the project) and I are incredibly proud of the whole team, and we’re so excited to be sharing this work and what long context and in-context learning can mean for you today! There’s lots of material about this, some of which are linked to below. Main blog post: blog.google/technology/ai/… Technical report: “Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context” goo.gle/GeminiV1-5 Videos of interactions with the model that highlight its long context abilities: Understanding the three.js codebase: youtube.com/watch?v=SSnsmq… Analyzing a 45 minute Buster Keaton movie: youtube.com/watch?v=wa0MT8… Apollo 11 transcript interaction: youtube.com/watch?v=LHKL_2… Starting today, we’re offering a limited preview of 1.5 Pro to developers and enterprise customers via AI Studio and Vertex AI. Read more about this on these blogs: Google for Developers blog: developers.googleblog.com/2024/02/gemini… Google Cloud blog: cloud.google.com/blog/products/… We’ll also introduce 1.5 Pro with a standard 128,000 token context window when the model is ready for a wider release. Coming soon, we plan to introduce pricing tiers that start at the standard 128,000 context window and scale up to 1 million tokens, as we improve the model. Early testers can try the 1 million token context window at no cost during the testing period. We’re excited to see what developer’s creativity unlocks with a very long context window. Let me walk you through the capabilities of the model and what I’m excited about!

@sampullara Spicy aioli

I feel like bacon, egg and cheese sandwiches should have a sauce.

@swyxio @samlambert sharding has to integrate too tightly with the database protocol & quirks for a generic solution to be as useful as you’d hope. (I worked on blog.twitter.com/engineering/en…)

This is why shared nothing architecture (sharding) works. A rogue traffic pattern is causing replica lag on a single shard and the problem is isolated. If this was a vertically scaled monolithic DB it would be a global issue.

@maccaw What level SWE do you expect to be displaced? L1 react? L9 systems?

I think we'll have automated software engineering by the end of the year.

My prediction is that by the end of 2024 we'll have practically solved hallucinations, models will be 30x as powerful, and AI will be better than humans at their jobs.

@antirez @Sheikheddy It has gotten worse

@Sheikheddy It fails even at trivial stuff like "write me a toy bloom filter". Don't even start trying to analyze some system programming code with GPT4 help. It does not understand anything basically, and allucinates random stuff.

The level of failure of LLMs at system programming tasks is incredible. They are almost useless, only good to provide memorized documentation. I guess system programming requires some more advanced reasoning? Because I don't believe it is under represented in training data.

@JacquesThibs @jeremyphoward HN’s algo heavily penalizes “ratioed” posts, which this no doubt was

@giansegato @GergelyOrosz There's a special clause that says building software is *always* R&D.

@GergelyOrosz one thing i don't get is why do you have to classify those costs as R&D? can't you just write them off as actual direct costs contributing to revenues? this would make them deductible no?

Turns out this IS a massive problem for so many US tech companies. There’s a general disbelief on how the government is ignoring all of it, still. Unless changed, the US tech sector becomes so much less competitive than the rest of the world for esp for companies in early stage.

@sampullara I wonder if the training away from hallucination (i.e. trying too hard) also trains away from effort.

This is what I have been seeing but I don't think it is new. It has always been lazy when the problem is large but I don't know why it would ever say something like this. Definitely need to make sure that it tries harder somehow.