CEO 'Fucking Hates' GenAI, text-to-game platform, & Models with Existential crises

TLDR;

Technology:

😈 Hermes 3, an LLM with Existential Crises

🔣 DeepSeek Prover-v1.5, a new LLM Math SOTA

👩‍🏫 Self-Taught LLMs, Meta’s New Evaluators

Markets:

🤝 The State of Enterprise GenAI by Deloitte. Problems persist.

🫢 Apheros’ new cooling technology improves status quo by 90%

Product:

🙃 I Really Fucking Hate Generative AI, CEO exclaims

🖥️ Text-to-game, models that turn text into videogames

LARGE LANGUAGE MODELS

Hermes 3, The Existential Crisis Model

Nous Research has released Hermes 3, a model fine-tuned from the Llama 3.1 models that boasts better performance in tasks like reasoning or function-calling (which are fundamental for agents) while being much less restricted than closed-source models, as seen in the image above.

But what has made the news is that, when provided an empty prompt, the model enters a sort of identity crisis state, with an example provided by the research lab, which is certainly weird.

This is important because…

On the one hand, Nous Research is doing God’s work by releasing uncensored LLMs that can be tailored to our particular needs and not stick to the prejudices and moral high ground that a few people back in Silicon Valley impose.

But I’m really not buying this weird behavior by the model that some have called ‘existential crisis.’ Trying to convey that an LLM has gained consciousness or ‘self-awareness’ seems utterly absurd and more like a PR move than anything else (you can fake these behaviors).

FRONTIER RESEARCH

DeepSeek-Prover-V1.5, a New Maths SOTA

Researchers from DeepSeek, an AI lab that has been on a tear lately, have announced a new math prover LLM that becomes the state-of-the-art for theorem proving in some high-school and undergraduate benchmarks.

The model combines an LLM with a new approximation of Monte Carlo Tree Search, or MCTS (similar to AgentQ we saw last Sunday), which incentivizes the model to search the solution space and increases its likelihood of success.

This is important because…

Combining LLMs and MCTS is becoming the norm. As we saw in our Thursday newsletter, rote multiple sampling (generating various possible solutions to a problem) yields better outcomes.

However, MCTS takes it to another level by allowing the model to sample multiple solutions and explore them with better intuition. In this case, just like AgentQ explored different ways to navigate a website, here Deepseek-prover explores different ways to solve a mathematics theorem.

You may have realized it sounds similar to Google’s AlphaProof, which we covered a few weeks ago. But before you ask, no, I don’t know which option is best, as I haven’t seen both approaches on the same benchmark.

But I can tell you that maths and coding are about to see a considerable increase in performance thanks to this combination of LLMs and MCTS for one simple reason: these modalities can use automatic verifiers that lead the search process (unlike in natural language tasks, where the verification is mainly subjective).

ALIGNMENT RESEARCH

Meta’s Self-Taught Evaluators

Meta researchers have proposed a new method that allows reward models…

Enjoyed the introduction? Subscribe or sign in to continue with this edition.