Entryism Enjoyer🐵🧦🏗️

So it appears the scandal involving Troy Jackson that was getting hyped up last week is that he and his longtime partner are second cousins. Marrying cousins is a common thing in Aroostook County where Jackson lives.

Formally verified compressible hydrodynamics in a complex geometry: simulating supersonic flow, shock wave interactions, and turbulent structure formation with the isothermal and full compressible Euler equations, all implemented and verified autonomously with Lanyon. ~7,000 lines of C, ~11,000 lines of Lean proof, ~150 correctness theorems, ~3 minutes. Euler's equations form the basis of our understanding of compressible hydrodynamics, are critical for aerospace and other engineering applications, and represent the underlying hyperbolic equation structure of more complex systems such as the Navier-Stokes equations and the ideal magnetohydrodynamics equations. Solving them correctly relies not only upon mathematical and numerical correctness, but physical and thermodynamic consistency too. To the best of our knowledge, Lanyon just one-shotted the first ever end-to-end formally verified solver for Euler's equations, complete with full and executable mathematical, physical, and numerical correctness proofs. And, for an encore, it ran the resulting simulations in complex (and even dynamic!) geometries. GitHub link and technical deep dive below. 👇

Also what IIT Madras did different from other IITs was good marketing around in-house deep-tech companies. Ather or ePlane also got good momentum because IIT Madras ecosystem reminded committed to their success. While IITs also created core tech startups like Baaz, Nocca, Vecmocon, the buzz remained largely around Unicorns.

*"SITUATION" NOT AS BAD AS FEARED--SOURCES *SALP -35% MTD, STILL UP SUBSTANTIALLY YTD *ANTHROPIC IS NOW 50% OF $20B REMAINING AUM *FIRM WAS 4X LEVERED AT TIMES *GSCO MET WITH SA TO DISCUSS ORDERLY LIQUIDATION LAST WEEK *SITUATION STILL FLUID (STACK) Godel Terminal

There are so many jihadists swarming the Sahel region of Africa that (a) most towns are building walls of berms and trenches and (b) some towns have been abandoned. Decline and fall vibes.

📢 𝐉𝐔𝐒𝐓 𝐈𝐍: $INTC Startup CEO Is Intel CEO's Co-Investor in Other Firms, Intel granted startup RosaicLabs access to certain technologies; RosaicLabs' CEO is a co-investor with Intel's CEO in another company.

Some napkin math on the K3 posttraining cost: - $4M total for the hero run - <500 steps per expert - 160K tasks/ domain ---- With 50M training rollouts and 24 experts (8 domains x 3 reasoning efforts), it comes to ~2M rollouts per expert. With 1.3M tasks, ~160K tasks/domain. But, suppose 80% of the 2M rollouts is RL and 20% of that is OPD, and assume group size 16, its closer to 100,000 tasks per expert. So lets just say 120K tasks. Looking at the RL plots, the jaggedness makes me think that the x axis isnt all that large which means fewer steps and larger batch size. Assume racks of GB300 NVL72 in their colocated setup. I dont know the training requirements for something this big so this is pure spitballing for the sake of getting something out. Each train replica can probably do 1024 batch size with KDA/MLA (if i had to guess it probably cannot be larger than 1024). They say each experiment is within a few hundred GPUs so maybe 8x72. 8 training replicas x 1024 batch size = global batch size 8192. 8192/group size 16 = 512 unique problems per batch. This comes to just under 250 steps, which seems ~reasonable based on the shape. Now, each training step probably takes something like 10 minutes, especially with colocation. So tldr per expert its ~2days on 576 B300s. Lets just say they have 2K gpus and can train 4 experts at once so expert training is 12 days in total. To simplify for OPD, lets just say 4 days so 16 days total. Suppose 1 B300 is 4.50/h, in total the whole thing costs ~$4M

GPT-5 was the best model in the world a year ago. Today it's inferior to Qwen3.6-27B, an open weight model you can run on your laptop. Progress is brutal.

After deployment, we applied GPT-5.6 Sol to advance the frontier of efficiency by making itself more efficient to run. The results: - 20% lower serving costs from production GPU kernel improvements. - 15%+ better token-generation efficiency from improved speculative decoding.

Introducing Orchid, the first assistant that actually gets you 🌸