chris

I'm very impressed by the way Phind, despite being the tiniest of startups, has managed to keep up with the giants. Phind-70B beats GPT-4 Turbo at code generation, and runs 4x faster. There is definitely still room for startups in this game.

Will GPT-5.5 surpass top1? "16MB artifact size" will induce some parameter efficient but not inference efficient methods. This is also interesting, cuz it targets Kolmogorov complexity. (I'm not very interested in this setting tho)

once you make the nyc -> sf journey enough times, your brain learns to make peace with the 6 hour flight and stops resisting

Completely fascinating demo, a glimpse of the future for research, coding, and image/video gen. There are real tradeoffs to consider The $30M spent claim is real but “extreme specialization” is too. Basically an ASIC per model release, from the co-founders of Tenstorrent

On sports hedging. The sports insurance and re-insurance industry is big: the annual market is around $9 billion and is projected to double by 2030. There are a variety of insurance products including brand sponsorships, game cancellations, team/player performance, off player compensation, and more. We just announced a partnership with sports insurance broker Game Point Capital. Game point capital issues hundreds of millions in sports insurance per year. Their most popular product is team and player performance bonus insurance: sports teams often structure large payouts to coaches and players that get triggered if they achieve certain milestones (winning championship, making the playoffs, scoring records, etc.). The bill is often large and teams smooth out their finances by hedging it. Game Point hedged for two different teams against performance bonuses on basketball with Kalshi last week: 1. One is a hedge for a bonus if the team makes the post-season (Kalshi price=6%, OTC price ~ 12-13%) 2. One is a hedge for a bonus if the team advances to the second round (price= 2%, OTC price ~7-8%). Why did they do this on Kalshi? Insurers like Game Point need to offload the risk they take on somewhere else. Typically, they go to traditional re-insurance companies like Lloyd's of London, which is Over-the-counter (OTC): you negotiate price and terms 1:1 with them (instead of an open/competitive market). Like in all OTC markets, the issue is the re-insurers are restrictive in what risks they take: they like to avoid volatile, higher risk contracts, so they offer prices that are opaque and prohibitively high. Exchanges are a better alternative because they expand liquidity and bring competition: multiple counterparties compete in an open marketplace to improve the price. Exchanges are harder to build than OTC because they need to have enough liquidity. Over the past year, we’ve massively increased the liquidity on our sports markets. During the Super Bowl, Kalshi could have processed a $22 million trade without moving the price meaningfully. At this level of liquidity, Kalshi is now very attractive for Game Point and other similar companies: there’s more liquidity available, it's cheaper, and the price is more transparent. We expect to process tens of millions in similar hedges from Game Point alone in the coming months. Onwards.

I also updated our Composer 1.5 blog post with Terminal Bench 2.0 scores since that was a common request. Pretty happy with our performance (better than Sonnet!) and yet still much more work to do. We're gonna need more GPUs 😄

We trained the best coding model in the world under 1T parameters. Excited for people to try it!