Sampling of feedback on Bitcoin from people I met from around the world during my travels in Europe this June: “I won’t touch that, it’s too volatile.” “Isn’t it the same price now that it was 5 years ago? What kind of investment is that?” “Something is holding it back from becoming a real money, what do you think it is?” “I like something I can touch, like gold.” “What do you think of the guy who sails who bought all the Bitcoin?”

Intel is sampling a new inference card. The pitch is capacity and power, not training speed. At the OCP Global Summit 2026 it started customer sampling of the GPU D-Series, Crescent Island. The D270 delivers 1.3 petaflops of FP8 and 1.5 terabytes a second of bandwidth in a 350-watt air-cooled PCIe card. Its own design has 160 gigabytes of memory. Partner designs reach 480 gigabytes of LPDDR5X. It is built on Xe3P for agentic AI inference and long-context work in ordinary racks. One card supports models up to about 300 billion parameters. Eight cards go past one trillion. Open tools like vLLM are supported. Sampling is underway with server makers and software partners. This is its push for high-capacity, lower-power inference next to Xeon, not a training rival. If one 350-watt card holds a 300-billion model, does that steal inference slots from Nvidia, or does it only sell where the rack cannot take a hotter GPU? $INTC

"Finetuning with Sampling: SFT Learns Better Than You Think" SFT may be better than we think, and the problem could have just been the data distribution. This paper uses MCMC to rewrite expert traces into equally correct answers that look more like what the base model would naturally produce. With that, plain SFT can outperform RL while forgetting less!

Standard inference time sampling methods (rejection sampling, best-of-n) can fail catastrophically as errors compound over long generations - even with a nearly-perfect verifier. @AShettyV talks through VGB (value-guided backtracking): a provably more robust fix, rooted in sampling theory. Watch the full tech talk here: