Naisbitt once said “We're drowning in information but starved for knowledge.” 58% now suffer fm info overload -- so Annotote’s knowledge network gives highlights of everything u need to read and lets u annotate anything u want to save or share. Do it >>> annotote.launchrock.com

Why did he get smaller between 13 and 23 years old…?

Answering straight up, without nuance or qualification (e.g. Continuous Learning #CL would completely blur line between training and inference)… 1️⃣ What percentage of current AI cloud revenue and semi demand is lab training? Training 30~50% of current AI cloud revenue/spend (and 40~60%) of leading-edge GPU demand rn. Inference is majority of ongoing workloads and is growing faster, ~½ of AI compute in 2025 and ~⅔ in 2026. 2️⃣ What percent of forward views of demand are driven by lab training? Training demand consensus declining rapidly, with 20~40% 2026-27e growth, falling LDD/HSD thereafter, because Ship of Theseus model improvements elongate major new model release cadence (e.g. fine-tuning/post-training/RAG on smaller cheaper runs rather than training 100B parameter models from scratch). Inference estimates 70~90% of total AI compute by 2030. 3️⃣ How much of this training demand is required if the future is enterprises doing post-training on open source models? Widespread open source adoption could cut overall training compute ~50% maybe, but already 76% of enterprise LLM users leverage some open source, which is constrained rn (e.g. data quality, talent, final-mile optimization). 4️⃣ How much inference capacity is unlocked if all this old training capacity is repurposed for inference? Substantial, maybe 30~50% of current high-end GPU capacity, already reflected in most analysts' capex models. 6️⃣ How do inference demand projections change if on-device inference takes off (like Apple bulls think with new Siri AI)? Edge inference ~70% share in certain segments, if successful at scale (privacy, latency, zero marginal cloud cost) – offsets some hyperscaler revenue but doesn't kill enterprise/cloud use cases needing massive context or heavy compute. 7️⃣ Is GPU, CPU, and memory demand constant across workloads? How do they vary? Training is GPU-heavy (high parallel FLOPS/HBM memory) and lower CPU. Inference varies, with GPUs dominant; CPUs increasingly prominent for orchestration, preprocessing, agentic, RL; and memory critical for KV cache (DRAM/HBM) and storage (NAND/SSD). Agentic and enterprise workloads usually increase CPU:GPU ratios in favor of CPU. 8️⃣ How important is latency for the majority of future enterprise workloads (low if agentic)? Latency matters but usually secondary for many enterprise/agentic cases. More here: x.com/i/status/2054976988775…. . . . 1️⃣0️⃣ Are memory requirements for NAND and SSD also constant? How do they vary? Training needs massive storage for datasets/checkpoints. Inference varies – high for large context windows/KV cache (but often in DRAM); lower for cached/static models; RAG/vector DBs boost SSD/NAND demand; while enterprise on-prem may increase local storage needs vs cloud. 1️⃣1️⃣ What % of future enterprise open-source workloads will be on-prem vs cloud? How does that compare vs today? Hybrid, with on-prem share 30~60% (e.g. sensitive/regulated workloads).

Google gonna recycle every redeemable device every human has ever owned, then harvest it for compute and memory – no more buying back your user cell phones, smartphones, and computers to refurbish and resell them lol, cuz now they AI clusters… $googl

"ammm, sir, why do you have 10 cell phones...?!"

"Richie's" has no right to be such a hard ass, given he's the heir of a slush empire…

See, that's not Landon Donovan, that's @bchesky…

The Musketeers¹ showed up – SpaceX $spcx got its IPO pop and Elon's the world's first trillionaire… ¹x.com/i/status/1526278088648… #musketeers #elon

Can somebody please explain to me how a gas station charging $4.25 ( 15% premium) can survive next to another gas station that charges $3.65…? Fwiw, it's almost always the Exxon Mobil $xom that's most expensive vs local entrepreneur that's price undercutting – which is another fascinating economies (diseconomies) of scale question…

Hitting a round-ball with a round-bat (baseball) might be the most difficult act in all of the major American sports, but hockey players are the most talented in aggregate – not even particularly close imho: sure, linebackers (football) and 7-footers (basketball) are freak athletes; but what the NHL does for 84 games, on skates, on ice, stickhandling an erratically prone disk, in traffic, at high speed, with opponents hanging all over them, while trying to knock-your-block-off… is a miracle on the scale of "it's unbelievable that humankind actually achieved this"… #skill