AI research wiki
Friday, May 15, 2026 · social stream

Media Live

daily roll-up

Summary

The strongest cross-slot cluster is xAI's Grok Build CLI launch in the morning (five tweets across @xai, @JasonBud, @milichab), the fourth-frontier-lab entry in the agent CLI category and a direct pair with today's WildClawBench harness-as-load-bearing finding. The single-slot standout is the evening's Anthropic CFO podcast tease from @nottombrown, with run-rate revenue stats ($9B to $30B in one quarter, NDR over 500%, 90% of internal code written by Claude Code) that, if half-true, are the most aggressive enterprise numbers ever publicly attributed to an AI lab. Secondary signal: @ClaudeDevs documents a concrete prompt-cache pre-warming trick (system prompt before user prompt) for time-to-first-token wins, and AWS announces a $110M Build on Trainium program targeting Berkeley/MIT/CMU researchers, replaying NVIDIA's CUDA-subsidy playbook on Neuron. Everything else is noise: Lex Fridman hitchhiking in China, NVIDIA commencement video, Cybertruck FSD marketing, two opaque x.com/i/article reposts, political content. Three slots ingested (morning, afternoon, evening); no night slot today.

Posts

  • xAI Grok Build CLI for SuperGrok Heavy (cluster of 5) (@xai · @JasonBud demo · @JasonBud subagents · @milichab /imagine · @milichab plan mode · x.ai/cli) [morning]. Early-beta agentic CLI with subagents, plan mode, clickable interface, and /imagine plus /imagine-video in-terminal. Five-lab race in agent CLIs now (Claude Code, OpenAI Codex, Hermes, Grok Build, plus internal Anthropic stack); pair with WildClawBench and treat harness as a first-class routing variable.
  • Anthropic CFO podcast tease: $9B to $30B run-rate, 90% internal code via Claude Code (cluster of 2) (@nottombrown intro · @nottombrown stats) [evening]. Krishna Rao's first podcast: NDR over 500% annualized, first revenue in March 2023, on pace for $50B run-rate next month, ~$75B raised, Cowork outpacing early Claude Code adoption. Reinforces the Anthropic-overtakes-OpenAI B2B thread; the Trainium/TPU/GPU allocation discussion is routing-adjacent infrastructure signal.
  • Anthropic prompt-cache pre-warming pattern (cluster of 2) (@ClaudeDevs · @ClaudeDevs docs · docs) [afternoon]. Send the system prompt before the user prompt so Claude writes it to cache without generating output, then the real request lands on a warm cache. Practitioner KV-cache adjacency for long-prompt API workloads, fits the kv-cache thread.
  • Position paper: agentic AI as the only foreseeable route to AGI (@omarsar0 via @bayesiansapien · arxiv 2605.12966) [afternoon]. Argues monolithic scaling is insufficient and agent DAGs achieve exponentially better generalization and sample efficiency, with explicit connections to MoE and routing topologies. Tier-2 read for the agentic-systems direction.
  • AWS $110M Build on Trainium program for university researchers (@mattsgarman · aboutamazon.com) [evening]. Berkeley, MIT, CMU and others get dedicated Trainium access, open-source contributions flow back to Neuron. NVIDIA-style CUDA-subsidy playbook for the non-NVIDIA stack; sits next to Amazon-Anthropic capital concentration and the broader gpu-kernels GPU-alternative push.
  • HBR: "AI brain fry" hits high performers hardest (@rohanpaul_ai via @bayesiansapien) [afternoon]. Survey of 1,500 workers reports AI intensifies workloads, forcing constant task-switching. Opinion/HR-research framing, soft responsible-AI signal on deployment-side cognitive load.
  • JasonBud calibration tweet on Grok Build (@JasonBud) [morning]. Sets early-beta expectations; surrounding cluster covers the substance. Skim.
  • "Vibes are hard to replicate in silicon" (@MillionInt) [evening]. One-liner reflection, no data. Skim.
  • Opaque x.com/i/article reposts (@ashwingop) [afternoon]. Article body not fetched; click through if curious.
  • NVIDIA Jensen Huang CMU commencement (@nvidia) [morning]. PR/motivational. Skip.
  • Lex Fridman hitchhiking-in-China announcement (cluster of 2) (@lexfridman trip plan · @lexfridman form) [morning]. Off-topic except as a downstream podcast signal on China AI engineers. Skip until episodes drop.
  • DAIR.AI Academy "Vibe Coding AI Apps with Claude Code" course (academy.dair.ai) [afternoon]. Course-catalogue side article. Skip.
  • Magicsilicon: Intel x McLaren F1 compute partnership (@magicsilicon) [morning]. Sponsorship PR. Skip.
  • Tesla Cybertruck FSD marketing (@Tesla) [morning]. Demo-drive promo. Skip.
  • @BrettRatner AF1/Starlink/Jensen photo (@BrettRatner) [afternoon]. Celebrity tweet. Skip.
  • WHFraudTF political posts (cluster of 2) (article · livestream) [afternoon]. Off-topic political content. Skip.
slot detail

Evening

scraped 2026-05-15 22:00 IST · 4 tweets

Summary

Tiny slot, but two genuinely consequential signals. @nottombrown (Anthropic) shills the Krishna Rao CFO podcast with a list of eye-watering numbers (cluster of 2): run-rate revenue $9B to $30B in one quarter, reportedly on pace for $50B by next month, NDR over 500% annualized, and over 90% of internal Anthropic code now written by Claude Code. AWS CEO @mattsgarman announces the $110M Build on Trainium program funding university researchers (Berkeley, MIT, CMU) on Trainium/Neuron, a clear hardware-stack play to widen the non-NVIDIA developer pool. @MillionInt offers a one-liner musing on how hard "vibes" are to encode in silicon, philosophical filler.

Posts

  • Anthropic CFO podcast: $9B to $30B in one quarter, 90% of internal code is Claude Code (cluster of 2) (@nottombrown · @nottombrown intro). Krishna Rao's first podcast appearance. Stated stats: NDR over 500% annualized, first dollar of revenue in March 2023, run-rate $9B to $30B in one quarter with $50B trajectory by next month, ~$75B raised, Cowork growing faster than Claude Code did at the same point, head of tax is the heaviest token user on the finance team. If even half-true these are the most aggressive enterprise numbers ever publicly attributed to an AI lab and reinforce the Anthropic-overtakes-OpenAI B2B thread. The Trainium/TPU/GPU compute allocation discussion is also a routing-adjacent infrastructure signal worth a listen.
  • AWS $110M Build on Trainium program for university AI researchers (@mattsgarman · aboutamazon.com). Berkeley, MIT, CMU and others get dedicated Trainium access; all research is open-source so kernel and compiler improvements flow back to the broader Neuron ecosystem. Same playbook NVIDIA ran with CUDA in the 2010s: subsidize the next cohort of researchers onto your stack so the tooling gap closes from the bottom up. Sits naturally next to the Amazon-Anthropic capital concentration thread and the broader GPU-alternative push covered in gpu-kernels.
  • "Vibes are hard to replicate in silicon" (@MillionInt). One-liner observation that humans reason well over uncertain nonquantifiable qualities and that this remains hard for models. Reflection, no data. Skim.

Afternoon

scraped 2026-05-15 15:00 IST · 8 tweets · 3 curated

Summary

Two genuinely useful items in this slot, both buried in a mostly noisy feed. @ClaudeDevs ships a concrete prompt-cache pre-warming tip (send the system prompt before the user prompt to cut time-to-first-token), and @omarsar0's repost surfaces a position paper arguing agentic AI, not monolithic scaling, is the only foreseeable route to AGI. A softer @rohanpaul_ai repost on "AI brain fry" from HBR is worth a mention for the responsible-AI angle but is opinion, not research. Everything else (a bayesiansapien repost behind an opaque x.com/i/article link, two political WHFraudTF posts, a BrettRatner photo tweet) is skip-tier.

Posts

  • Pre-warm the Anthropic prompt cache to cut time-to-first-token (cluster of 2) (@ClaudeDevs · @ClaudeDevs docs link · docs). Send the system prompt before the user prompt. Claude writes it to the cache, skips generating output, and the real user request lands on a warm cache. Cheap concrete latency win for long-prompt API workloads. Practitioner-grade KV cache adjacency, fits the kv-cache thread.
  • Position paper: agentic AI as the only foreseeable route to AGI (@omarsar0 via @bayesiansapien · arxiv 2605.12966). Authors argue scaling a single model is insufficient and that agentic systems achieve exponentially better generalization and sample efficiency, with theoretical derivations contrasting monolithic learners against agent DAGs. Explicit connection to Mixture-of-Experts and to the instability of current multi-agent frameworks. Tier-2 read for the agentic-systems direction, with a routing-flavored core (simple routers generalized to DAG topologies).
  • HBR: "AI brain fry" hits high performers hardest (@rohanpaul_ai via @bayesiansapien). Survey of 1,500 workers reports AI intensifies workloads rather than reducing them, forcing constant task-switching and oversight. Opinion / HR-research framing, not a measurement paper, but worth tracking as a responsible-AI signal on deployment-side cognitive load.
  • DAIR.AI Academy launches a "Vibe Coding AI Apps with Claude Code" course (academy.dair.ai). Surfaced as a side article in the omarsar0 thread. Course catalogue, $29/mo Pro tier. Skip.
  • Opaque x.com/i/article repost from @bayesiansapien (@ashwingop). Article body not fetched. Click through to read.
  • @BrettRatner photo of AF1 with Starlink and Jensen tag (@BrettRatner). Celebrity tweet. Skip.
  • WHFraudTF political posts on JD Vance Maine remarks (cluster of 2) (@WHFraudTF article · @WHFraudTF livestream). Off-topic political content. Skip.

Morning

scraped 2026-05-15 09:00 IST · 10 tweets

Summary

The morning slot is thin and almost entirely captured by xAI's Grok Build CLI launch. Five tweets across @JasonBud, @milichab, and the @xai mainline ship an "early beta agentic CLI for coding, building apps, and automating workflows" for SuperGrok Heavy subscribers, with /imagine and /imagine-video commands directly inside the terminal. No @bayesiansapien retweets in the lookback window. NVIDIA contributes one Jensen Huang CMU commencement video. Lex Fridman posts travel-recommendation forms (off-topic) and a hitchhiking-in-China announcement. Tesla pushes Cybertruck FSD marketing. Magicsilicon promotes Intel as McLaren F1 compute partner. The single signal worth click-through is the Grok Build CLI as the fourth-frontier-lab entry in the CLI agent category, directly pairing with today's WildClawBench finding that harness choice shifts performance by 18 points.

Posts

  • xAI ships Grok Build CLI for SuperGrok Heavy (cluster of 4) (@xai mainline · @JasonBud · @JasonBud subagents · @milichab on /imagine · @milichab on plan mode + syntax highlighting · x.ai/cli). Early beta agentic CLI. JasonBud demos subagents (the model checks its own work via subagents, "running 5+ agents at a time and jumping between plans"), clickable interface, plan mode. milichab demos /imagine and /imagine-video for in-terminal image and video generation. Direct fit for the harness-as-load-bearing thread from today's WildClawBench: the agent CLI category is now a five-lab race (Claude Code, OpenAI Codex, Anthropic Claude Code, Hermes Agent, Grok Build), and harness choice moves a model by up to 18 points. Cross-link to llm-routing.md — harness should be a first-class routing variable, not a deployment detail.
  • JasonBud calibration tweet on Grok Build (@JasonBud). Sets expectations: "this is an early beta. we still have much to improve." Useful disclosure framing; the goal is "Grok Build to be the best coding agent when it comes to solving real and difficult engineering problems." Skip body, the surrounding cluster covers the substance.
  • NVIDIA Jensen Huang CMU commencement (@nvidia). "No generation has entered the world with more powerful tools or greater opportunities." Speech to CMU class of 2026. PR / motivational. Skip.
  • Lex Fridman travel + hitchhiking-in-China announcement (cluster of 2) (@lexfridman travel form · @lexfridman trip plan). Lex on a backpacking trip starting in China, hitchhiking across rural China after meeting engineers at the heart of "China's AI revolution." Off-topic for the wiki except as a downstream Lex Fridman Podcast signal (China AI engineer interviews likely to follow). Skip until podcast episodes drop.
  • Magicsilicon: Intel × McLaren F1 compute partnership (@magicsilicon · Intel page). PR / sponsorship announcement. Intel marketing-side activity. Skip.
  • Tesla Cybertruck FSD marketing (@Tesla). Demo drive promotion. Skip.