New Method to Train AI Models 44% Faster, Overspending on Computing Infra, Sam Altman Praises Kosmos

Captions as YouTube publishes them, usually machine-generated, so product and people names can be misheard.

1. 0:00

   Good morning Legion. We are one day out from the release of the next set of missions. The missions needed to reach the Vanguard rank. We're also going to be awarding the first Legion members with their ascendant rank SBTs accompanying PFPs.

2. 0:13

   We're also in the process of playing the first IRL events for the Near Legion, set to kick off in January of next year. Ten locations around the world will be chosen. to host these events with modest budgets to cover pizza, drinks and venues.

3. 0:25

   If you're interested in hosting an event in your local area, let us know in the Legion chat as you prepare to select our first hosts. And now for the good stuff. Lots of AI news. Our first story up today is a new paper between Tencent and Tsinghua called Continuous Autoregressive Language Models or CALM The paper challenges the next token paradigm every LLM is built on.

4. 0:46

   To put it simply, the team is proposing a new way to train LLMs to think more efficiently and token effectively. Now, why does this matter? Because if the team can prove and scale this framework, it would lead to four times fewer prediction steps and 44% less training compute required in future LLMs.

5. 1:05

   As a reminder, AI is always changing. Just like the internet. Just like the evolution of crypto. The early winners are not always the long term winners. Our second story may be a sign of the AI bubble popping. The founder of Klarna, Sebastian, worries that the tech industry is committing too much money to huge compute infrastructure.

6. 1:24

   He tells the Financial Times that he believes that less computing power to run advanced AI was needed than tech companies were planning to spend, as the models themselves were the most cost effective compression technology ever invented, which, as he argues, will run more efficiently over time, meaning it will cost less over time, meaning less money will be needed for huge compute infrastructure over time.

7. 1:48

   Is this the canary in the coal mine? Is hyper financialization from the frontier models actually going to drive their progress backwards as Chinese models continue to improve on a much cheaper budget? Third, and finally, we've talked about Kosmos before, but since that time, Sam Altman and many others have come out praising the product as a cutting edge use case for AI.

8. 2:09

   What's important to note here is that Kosmos uses a structured world model to integrate a data analysis agent and a literature search agent, so two separate agents working together. It is closed source. And to the extent that we know not based on any other model that exists today, this world model enables it to process orders of magnitude more information than traditional language models can handle within their context limits, suggesting a bespoke design tailored for autonomous scientific discovery.

9. 2:37

   The most important thing to understand about Kosmos is that it was custom built with a very specific purpose in mind. It lives outside of the modern debate surrounding closed source frontier models versus open source eastern models.

10. 2:50

    It is, however, fully closed source, meaning the control outputs and capacity of cosmos will always be dependent on Edison Scientific acting in good faith for users looking to make their own scientific discoveries. That's all for today, folks.

11. 3:04

    Stick around in the next couple of weeks will be two NEAR AI product announcements that we're quite excited about.