Latent Space

appreciate your constructive feedback!

> i wonder if there are new ops solutions for the realtime apis popping up

This is something we have spent quite some time on already, both on designs internally and talking to teams using Langfuse with realtime applications. IMO the usage patterns are still developing and the data capturing/visualization needs across teams is not aligned. What matters: (1) capture streams, (2) for non-text provide timestamped transcript/labels, (3) capture the difference between user-time and api-level-time (e.g. when catching up on a stream after having categorized the input first).

We are excited to build support for this, if you or others have ideas or a wishlist, please add them to this thread: https://github.com/orgs/langfuse/discussions/4757

> retries for instructor like structured outputs mess up the traces, i wonder if they can be tracked and collapsible

Great feedback. Being able to retroactively downrank llm calls to be `debug` level in order to collapse/hide them by default would be interesting. Added thread for this here: https://github.com/orgs/langfuse/discussions/4758

> chatgpt canvas like "drafting" workflows are on the rise (https://www.latent.space/p/inference-fast-and-slow) and again its noisy to see in a chat flow

Can you share an example trace for this or open a thread on github? Would love to understand this in more detail as I have seen different trace-representations of it -- the best yet was a _git diff_ on a wrapper…

I had access to GitHub Copilot as a student in early 2022 while learning Haskell and immediately realised that it would hinder my learning if I didn't turn it off and implicitly follow this understand, plan, execute, reflect loop.

AI products like Cursor have the notion of an 'autonomy slider' [1] that can fortunately be turned all the way down (disable Cursor Tab) but relying on this discipline seems fickle when with the right agentic loops [2] and context engineering, thousands of lines of code can be churned out with minimal supervision.

I've considered always working on two projects over a long timespan, one with no AI assistance, possibly in a separate IDE like Zed, and one in Vibe Kanban (my current daily driver) but this feels like an inefficient proxy to accelerating this four step learning loop with a tool like solveit.

Since the solveit product isn't released and seemingly isn't competing with solutions, is there an opportunity to convey how AI product developers should be thinking about amplifying their users and keeping them in the learning loop?

So far, I've seen Claude Code's Learning output style [3], and also ChatGPT's study mode but in these cases, the only product change is a prompt and solveit is more than that.

[1] https://www.latent.space/i/166191505/part-a-autonomy-sliders [2] https://simonwillison.net/2025/Sep/30/designing-agentic-loops/ [3] https://docs.claude.com/en/docs/claude-code/output-styles#built-in-output-styles

Yes, there's a podcast with the post-training lead for L3 where he mentions this. Lemme try and find it.

edit: found it. The money quote is here, but I really recommend the entire podcast since it's full of great tidbits and insights.

> Thomas [00:33:44]: You mean between supervised fine-tuning like supervised fine-tuning annotation and preference annotation? Yeah. So 100% to RLHF. In fact, that's quite interesting. You start for Llama 2 with a pre-trained model and you have to have an instruction model to chat model. Otherwise, like the model is just like continue finishing sentences. So you need that to start RLHF. So we had to annotate like 10,000 examples. What did we do for Llama 3? You start with a new pre-trained model and then you want, before starting the RLHF, to have now a chat model, which is not too bad. The option one was, let's do human annotation again, like SFT stage. But in fact, by the principle I said before, the annotation would be actually worse than Llama 2. So what we did is that we generated all the data on the prompts with Llama 2 and we applied like basically the last round of Llama 2 we had to kick off and start Llama 3 post-training. So Llama 3 post-training doesn't have any like human written answers there basically, almost. It's just leveraging pure synthetic data from Llama 2.

https://www.latent.space/p/llama-3

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

There was an interesting interview with David Luan about this recently. For context, he was a co-lead at Google Brain, early hire at OpenAI, and is now a founder at Adept: https://www.latent.space/p/adept

The TL;DR on his take is that there are organizational and cultural issues that prevent Google from focusing their research efforts in the way that is necessary for what he calls "big swings," like training GPT-3.

In regards to your second question, Google's reputation in ML is definitely not hype. Purely on the research side, Google has been behind some of the most important papers in modern ML, particularly around language model. The original Transformers paper, BERT, lots of work around neural machine translation, all of the work that DeepMind has done post-acquisition, and the list goes on. On the applied side, they also have some of the most successful/widely-adopted ML-powered products on the market (think RankBrain/anything involving a recommendation engine, Translate, Maps, a ton of functionality in Gmail, etc).

Do you want to train models from scratch, or do you want to build cool things on top of AI models?

If the former, I suggest digging into things like the excellent Fast AI course: https://course.fast.ai/

If the latter, the (relatively new) keyword you are looking for is likely "AI Engineer" - https://www.latent.space/p/ai-engineer

There's an argument that deep knowledge of how to train models isn't actually that useful when working with generative AI (LLMs etc) - knowing how to train or fine-tune a new model is less useful that developing knowledge of the other weird things you have to figure out about prompting, evals and using these models to build production-quality apps.