Mistral was quietly watching everyone write them off in the AI race and decided it was time to drop their own pro chat model. You can try it for free. They also announced a partnership with Helsing to develop defence focused foundation models. Paris is in full fibrillation today for the AI summit, so new stuff could come out at any time.

I've also been using a Chinese one that dropped this week called Kimi. Starting to get hard to keep track of them all.

Some insane stuff on video as well with Goku.

You really can't sleep soundly if you're sama. Elon is bidding almost $100B for OpenAI, in a move that seems more of a wrench in the wheels than an interested attempt. Good explanation here.

A pretty good counter to the main narrative that hardware is not needed. Long read but summary here: 
- The future of AI belongs to those who can keep innovating novel topologies.
- The speed of innovation is determined by the rate of experimentation.
- The rate of experimentation is compute-bound.
- Deploying more hardware is hitting resource limits at global scale & developing better hardware is hitting dimension limits at atomic scale.

Pretty pessimistic take on VC in the AI era.

TL;DR

AI investing is closer to deep tech than Internet/SaaS investing over the past 20 years. The shift to reasoning-heavy AI models introduces significant operating expenses at scale, challenging the traditional software valuation playbook. While higher revenues may offset some costs as agents replace human labor, startups face massive challenges: higher opex, lower margins, and intense competition from incumbents. VCs (we) need to adapt or risk suboptimal returns. But considering the amount already invested, it’s likely the returns are already in the post.

I tend to not see it as dark, but it's hard to go against each point made here.

Gumroad is no longer hiring junior or even mid-level software engineers.

New process:

1. Sit and chat about what we need to build, doing research with Deep Research as we go.
2. Have AI record everything and turn it into a spec. 
3. Clean up the spec, adding any design requirements / other nuances.
4. Have Devin code it up.
5. QA, merge, (auto-)deploy to prod.

A lot of people will thrive in this new world, but to think that a lot of mid CS graduates won't have work is quite incredible.

Time to value in AI is honestly hard to grasp.

These use cases show just how much traditional software companies have to gain from implementing already available AI without having to develop anything revolutionary internally.

Code and weight released for the first general-purpose robotic foundation model that can be fine-tuned to a diverse range of tasks (folding laundry, cleaning a table, scooping coffee beans, etc) and that can control a variety of different robot type (robots with two arms, single-arm platforms, mobile robots, etc).
You can check it out in Hugging Face as well.

- Do robotic foundation models make sense?
- How should we think about vertical robotic solutions?
- How important are LLMs for robotics?
- How far can the sim-to-real gap be closed?
- Does form factor matter?
- What do customers care about?
- Do you need to be an end-to-end solution provider?
- Sense, think, act – what matters most?
- Unbundling or rebundling of the stack?
- Are there any novel business models that make sense?

Worth a read

One of the big news of the week. Figure is going at it by themselves on the AI front. At the same time, OpenAI didn't waste any time and is looking for people to build robotic hardware and filed a trademark for it as well.

The race is on!

Graphene keeps on surprising.

The surprise here lies not only in the occurrence of superconductivity itself, but also in evidence that these systems exhibit a “strange metal” phase, which is often linked to high-temperature superconductivity in other materials. This suggests that the interactions between electrons in graphene are far more nuanced than expected, providing a simpler platform to test theories originally developed for more complex superconductors.

"Generation 4 Bio AI: As we integrate advanced ML models, microfluidics, and robotics into our experimental workflows, we edge closer to a future where autonomous science doesn’t just support human-led research but transforms it and pushes it into new modalities and use-cases."

Classic science work, and hard to actually grasp how it could be used, but nonetheless wild. Photonics is clearly having a moment, and it seems clear that AI will push for more and more photonics chips, to get gains in energy efficiency and real-time data processing.

Sort of goes like this:
- cold therapy can kill cancer cells
- not everyone can do cold therapy
- so let's liposuct the fat cells
- CRISPR them to repogram them to use more calories
- put them back in the body

Seems like something that'd be useful even in non-cancer patients, essentially turning beige fat into brown fat

We've seen and cited quite a few quantum teleportation papers, but now scientists at Oxford seem to have been able to use teleportation on full quantum logical gates. 

And a short reminder by Paul Graham that you're not building for the hype. Just keep building, the future will come. 

From one of our portfolio founders! Europe's rigid patent system stifles creativity and collaboration. Josef advocates for reforms similar to those in the US and China, proposing a provisional patent system and a grace period to empower inventors and researchers.

Does consciousness arise from neural processes or is it a fundamental aspect of reality itself? Can we understand the universe just thanks to consciousness?

In case someone missed it, we're one year into the telepathy age. Neuralink made a good summary with the stories of Noland, Alex, and Brad, the first (?) three humans with telepathy. 

Sifted is showcasing our friends at 50 Years as they expand to Europe bringing their 5050 program to various universities ultimately trying to get many more young academics to launch startups.

Pretty big deal: the successful synchronization of the Baltic states' electrical systems with the Continental European grid marks a significant milestone in energy independence, effectively severing ties with the Russian-controlled networks.

Pretty unexplainable move from A16Z, other than attracting attention and wanting to be seen/positioned as extreme right in the market. Mr. Penny, who was acquitted after choking a mentally ill subway passenger to death, was hired to join the firm’s American Dynamism investment team with no previous experience. Let's see what this yields.

Payload estimates SpaceX's revenue to have reached $13.1B in 2024, up from $8.7B in 2023.

And they're not even on Mars. 

Cool new science-focused film company.

Epic progress on the Vesuvius Challenge (trying to read ancient scrolls from Pompei with AI).
Really exciting, and just goes to show you can just do things. Researchers are fermenting with the possibility that they'll have the biggest influx of data in their field since the Renaissance.

Compute, wet labs, resources, and a $125k/yr stipend to work on how AI will change biology and chemistry. Deadline on February 14.

Pretty damning piece on the Economist about universities, saying they "too often they generate ideas that no one knows how to use"