Latent Space: The AI Engineer Podcast

fun story[0] about Bard's June update - implicit code execution (remember? https://news.ycombinator.com/item?id=36229782) - the example that they advertised on their blogpost/video demo now no longer works.

Here's what I got as of today: https://g.co/bard/share/2580222b6fb2 (screenshot from a couple days ago: https://twitter.com/swyx/status/1678495067663925248?s=20)

> What are the prime factors of 15683615?

> (after outputting some code) The output of the code is the list [5, 11, 23, 43]

this result is hallucinated - Bard tells you it is running the code, but it doesnt. If you run the code, you get [5, 151, 20773], which actually is the correct answer

now lets all acknowledge that maintaining stable performance of LLMs is an unimaginably hard problem, but* i suggest some regression testing to keep up the baseline. Would fuzz the test (really, its just another eval) because of course this test is going to be memorized in the weights soon

0: i first discussed this on monday's update https://www.latent.space/p/code-interpreter#%C2%A7misc-notes...

* Edited based on feedback from commenters, thanks.

I think you are asking specifically about practical LLM engineering and not the underlying science.

Honestly this is all moving so fast you can do well by reading the news, following a few reddits/substacks, and skimming the prompt engineering papers as they come out every week (!).

https://www.latent.space/p/ai-engineer provides an early manifesto for this nascent layer of the stack.

Zvi writes a good roundup (though he is concerned mostly with alignment so skip if you don’t like that angle): https://thezvi.substack.com/p/ai-18-the-great-debate-debates

Simon W has some good writeups too: https://simonwillison.net/

I strongly recommend playing with the OpenAI APIs and working with langchain in a Colab notebook to get a feel for how these all fit together. Also, the tools here are incredibly simple and easy to understand (very new) so looking at, say, https://github.com/minimaxir/simpleaichat/tree/main/simpleai... or https://github.com/smol-ai/developer and digging in to the prompts, what goes in system vs assistant roles, how you guide the LLM, etc.

(author of https://www.latent.space/p/ai-engineer here)

this is a fantastic followup post spelling out the toolkit of the AI Engineer. I have a version of this in my notes but didnt want to publish it for fear of being too biased by ommission, plus it's generally good to leave others to fill in the blanks on some things. the Projects list is particularly great - one could build a course around these few things and be reasonably confident that you have the base skills of a competent "AI Engineer" (that, for example, any PM or founder could ask to do an AI project and they'd be well equipped to advise/make technical decisions).

don't miss Andrej Karpathy + Jim Fan's take on the AI Engineer role: https://twitter.com/karpathy/status/1674873002314563584

for those who dont like Twitter, i cleaned up the audio of our AI Engineer conversation with Jared Palmer, Alex Graveley, Joseph Nelson, and other self identifying AI Engineers here: https://mixtape.swyx.io/episodes/the-rise-of-the-ai-engineer...

it's a fun time for those who want to carve out a space for themselves specializing in this stuff.

You might be disappointed.. It seems to have been weakly confirmed by Geohot that this sort of mixed system is already what GPT-4's 'secret sauce' is. [1] It's something I've also been speculating for months on. Ctrl+F for "220 billion in each". His phrasing, numbers, and details are suggestive of a leak unless he's just completely blowing smoke - and I don't think there's any real reason to think he is.

It looks like they were pumping the model size up, started getting diminishing returns, and so turned to a mixed model of 8 expert systems meet LLMs - to try to keep squeezing out just a bit more juice. I think Geohot offered an incredibly insightful quote, "...whenever a company is secretive, it's because they're hiding something that's not that cool. And people have this wrong idea over and over again that they think they're hiding it because it's really cool."

It also goes a long way to explain their recent, ultimately unsuccessful, gambit in congress. If they don't see any way of overcoming the diminishing returns, then they're like the guy who got a headstart in a race where distance = time ^ (1/2).

[1] - https://www.latent.space/p/geohot#details

> Challenges in long-term planning and task decomposition: Planning over a lengthy history and effectively exploring the solution space remain challenging. LLMs struggle to adjust plans when faced with unexpected errors, making them less robust compared to humans who learn from trial and error.

While working on smol-developer I also eventually landed on the importance of planning as mentioned in my Agents writeup https://www.latent.space/p/agents . I feel some hesitation with suggesting this because it suggests I'm not deep-learning-pilled, but I really wonder how far next-token-prediction can go with planning. When I think about planning I think about mapping out a possibility space, identifying trees of dependencies, assigning priorities, and then solving for some kind of weighted shortest path. That's an awful lot of work to expect of a next token predictor (goodness knows its scaled far beyond what anyone thought - is there any limit to next token prediction?).

If there were one focus area for GPT-5, my money would be on a better architecture capable of planning.

(OP here) - yeah i know, but i also know how AI twitter works so I put both the headline and the caveats. i always hope to elevate the level of discourse by raising the relevant facts to those at my level/a little bit behind me in terms of understanding. think theres always a fine balance between getting deep/technical/precise and getting attention and you have to thread the needle in a way that feels authentic to you to do this "job"

ultimately my goal is to Learn in Public and demonstrate to experts that spending their time teaching/sharing with me is a good use of time because i will augment/amplify/simplify their message.

(pls give the full podcast a listen/read/watch, George went deep on tinygrad/tinybox/tinycorp and there's lots there he IS the authority on, and people are overly fixated on the GPT4 rumor https://www.latent.space/p/geohot#details )