artificial-intelligence

Multi-context source (vault/sources/) — also tagged for projects/stock-market and threads/politics. Speaker-aware (AssemblyAI; named: Jason Calacanis, Chamath Palihapitiya, David Friedberg, Gavin Baker — Sacks out this episode). All-In E274 — the AI-thread-side concerns Karpathy joining Anthropic, the EBIT-positive Anthropic milestone, and the recursive-self-improvement framing.

• Updated andrej-karpathy (person, +1 quote; physical home stays in vault/threads/artificial-intelligence/wiki/entities/): Chamath's framing of Karpathy's track record — "wave upon wave" of AI, FSD bitter-lesson commercialization, recursive self-learning puts models on "overdrive and autopilot."
• Updated anthropic (entity, physical home vault/threads/artificial-intelligence/wiki/entities/): added "Profitability + Karpathy hire (May 22, 2026)" subsection — EBIT-positive milestone per WSJ; $100B combined OpenAI+Anthropic ARR at 80% gross margins on inference (settles "AI ROI" bear case); 3-6 month lead over peers, 6-12 months ahead of open source; Karpathy joining to lead recursive-self-improvement pre-training team.
• Updated autoresearch-recursive-self-improvement (concept): added "Karpathy joining Anthropic to lead recursive self-improvement (May 22, 2026)" subsection — this is the operational consummation of the March 2026 personal-prototype framing at frontier-lab institutional scale. Chamath quantifies the rate as 10x/year ("new Moore's Law"); Baker pairs recursive-self-improvement with continual learning as "the two final frontiers." Falsification threshold sharpens — Anthropic capability disclosures over the next 12 months can be tracked against the 10x/year claim.
• Updated gavin-baker (cross-context — physical home: vault/projects/stock-market/wiki/entities/; tag thread/artificial-intelligence appended; +2 quotes deduped from stock-market ingest).
• Updated chamath-palihapitiya (cross-context — physical home: vault/projects/stock-market/wiki/entities/; quotes deduped from stock-market ingest).
• No new mechanism page — the recursive-self-improvement leg is captured by the existing concept page and now anchored at frontier-lab scale via Anthropic. If a multi-step causal chain emerges from Anthropic's next capability disclosure (e.g., Karpathy team → X capability gain → Y downstream effect → tradeable Z), promote then.

Multi-context source (vault/sources/) — also tagged for projects/stock-market. Speaker-aware (AssemblyAI; named: Dwarkesh Patel, Rainer Pope — name mis-spelled by AssemblyAI; wiki uses Reiner with Rainer alias). Dwarkesh blackboard-lecture episode walking through chip architecture from the bottom up.

• Updated reiner-pope (cross-context — physical home: vault/projects/stock-market/wiki/entities/; tag thread/artificial-intelligence appended; quote deduped from stock-market ingest).
• Updated dwarkesh-patel (cross-context — physical home: vault/projects/stock-market/wiki/entities/; thread/artificial-intelligence tag already present; quote deduped).
• No new concept pages — the lecture content is hardware-substrate explainer (logic gates, matrix mult, GPU/TPU/FPGA architecture), which serves as background for existing AI-thread concepts on capability/compute but isn't itself a claim-bearing concept worth filing as a standalone page. Future ingest passes could surface specific claims (e.g., "AI chips want matrix multiplication; multiply-accumulate is the binding primitive") onto existing concepts like llm-as-commodity-thesis if a follow-up episode connects the architecture to specific commodity-vs-moat economics.

Multi-context source (vault/sources/) — also tagged for projects/stock-market (ingested there first, on the markets/TSMC/Nvidia-moat/IPO lens). This AI-thread pass captures the AI-infrastructure + open-vs-closed-models angle and mostly hits the cross-context-reuse branch (the stock-market ingest already created the Cerebras/Feldman page cluster). Speaker-aware ingest with named speakers Joe Weisenthal, Tracy Alloway, Andrew Feldman; unnamed labels B/C/F/G/H/I are ad/clip voices → source-attributed only, no entities.

• Created open-vs-closed-source-model-economics (concept, status: active, conviction: medium): the AI-thread-specific synthesis the stock-market ingest didn't make. Open weights ~3–5% behind closed on quality but far cheaper per unit of intelligence (you pay serving cost not training cost); Joe's "no free AI software" + levelized-cost-of-intelligence framing; Tracy's quiet-enterprise-migration-to-Chinese-open-models (Kimi, Qwen) observation; Feldman's "no single winner" x86/ARM/custom analogy. Cross-linked to llm-as-commodity-thesis and cuda-moat-erosion-at-inference as the open-vs-closed pricing structure underneath model commoditization.
• Created joe-weisenthal (person, 4 quotes attributed across open-vs-closed-source-model-economics and inference-speed-as-a-pricing-premium): Odd Lots co-host; physical home is the AI thread (first context to record him as a speaker). Roles + alias @TheStalwart.
• Updated andrew-feldman (cross-context — physical home: vault/projects/stock-market/wiki/entities/, tag thread/artificial-intelligence appended, +4 AI-angle quotes via _lib/person_entity.py): the agentic-inference-speed rebuttal, Kimi-K2-serving open-economics claim, the 3–5% closed-quality gap, and the closed-vs-open "battle underway" framing. (Now 13 quotes total across two context lenses.)
• Updated tracy-alloway (cross-context — physical home: vault/projects/stock-market/wiki/entities/, tag thread/artificial-intelligence appended, +2 AI-angle quotes): the enterprise-migration-to-open observation and the "crux of the AI valuation argument" closed-vs-open / fast-vs-cheap premium framing.
• Updated cerebras (cross-context — physical home: vault/projects/stock-market/wiki/entities/, tag appended): new "Why it matters to artificial-intelligence" section (memory-bandwidth account of inference latency; open-weight serving cloud; agentic-inference-speed anchor); +2 Key facts (open-model cloud, speed-is-architectural); cross-link to open-vs-closed-source-model-economics.
• Updated nvidia (cross-context — physical home: vault/projects/stock-market/wiki/entities/, tag appended): new "Why it matters to artificial-intelligence" section reframing CUDA-moat decay as an ecosystem fact (which silicon frontier labs run on), cross-linked to the thread's demand-side inference-economics pages and open-vs-closed-source-model-economics.
• Updated g42 (cross-context — physical home: vault/projects/stock-market/wiki/entities/, tag appended): new "Why it matters to artificial-intelligence" section on the sovereign-AI deployment pattern (English-Arabic frontier models, genomic work). Thin evidence flagged.
• Updated inference-speed-as-a-pricing-premium (cross-context concept — physical home: vault/projects/stock-market/wiki/concepts/, tag appended): added "AI-thread angle" Evidence subsections — the answer-vs-agentic-inference distinction + Feldman's "dead wrong" rebuttal of Ben Thompson, and Joe's treadmill-of-expectations + inference-allocation-skill framing. Cross-linked to joe-weisenthal and open-vs-closed-source-model-economics.
• Updated cuda-moat-erosion-at-inference (cross-context concept — physical home: vault/projects/stock-market/wiki/concepts/, tag appended): added "AI-thread angle" Evidence subsection tying the 10-keystroke runtime portability to llm-as-commodity-thesis and framing the three-frontier-models-on-three-silicon-stacks fact as substrate diversification. Cross-links added.
• Updated inference-demand-to-wafer-scale-advantage (cross-context mechanism — physical home: vault/projects/stock-market/wiki/mechanisms/, tag thread/artificial-intelligence appended; no new steps — the chain is already complete and AI-relevant as-is).
• Updated cuda-moat-erosion-to-nvda-rerate (cross-context mechanism — physical home: vault/projects/stock-market/wiki/mechanisms/, tag appended; added one AI-thread non-tradeable reading to Implications: same chain read as AI-compute-substrate fragmentation + runtime portability removing inference-layer lock-in; related-concepts/Related extended with open-vs-closed-source-model-economics + llm-as-commodity-thesis). No new step — the stock-market chain already captures the spine.
• Updated index: added andrew-feldman, joe-weisenthal, tracy-alloway under People; new "AI infrastructure / silicon (cross-context with stock-market)" Entities subsection with cerebras, nvidia, g42; new "Compute and inference economics" Concepts subsection with open-vs-closed-source-model-economics + the two cross-context inference concepts; the two cross-context mechanisms added under Mechanisms; the source added under Sources with a multi-context annotation.

Mechanisms: no new mechanism page created. The two existing Cerebras mechanisms already capture this episode's causal spine; the open-vs-closed economics is a structural comparison better held as a concept than forced into a chain. Per step 5b, zero new mechanism pages is a valid outcome — chains were not forced.

Cross-context handoff: stock-market ingested this source first (markets lens: IPO/$64B cap, the constraint-routing supply thesis, the NVDA-moat-rerate read, sovereign-demand/CFIUS). This AI-thread ingest is independent and takes the AI-infrastructure + open-vs-closed-models lens; the shared person/entity/concept pages now carry both context tags.

Contradictions surfaced: 0. The hosts' speed-durability skepticism (already on inference-speed-as-a-pricing-premium as a both-sides tension) and the self-interested-CEO caveat on the open-economics claims are recorded as tensions on the relevant pages, not as queue-blocking contradictions.

Speaker-aware ingest of Dwarkesh Patel × Eric Jang (2h37m AlphaGo-from-scratch blackboard lecture). The thread's first deep RL-methodology source: AlphaGo's MCTS as a per-move supervised-target generator vs. LLM policy-gradient RL's credit-assignment problem, RL information inefficiency, and a first-person read on what AI research LLMs can/can't automate.

• Created mcts-vs-llm-rl-credit-assignment (concept, medium-high): MCTS supplies a strictly-better action per move (groundable value fn) → sidesteps credit assignment / low-variance supervised target → LLM RL credit-assigns one scalar reward across 100k+ tokens → sample-inefficient. ## The chain summary points at the canonical mechanism.
• Created rl-information-inefficiency (concept, medium-high): Dwarkesh's bits-per-flop = samples-per-flop × bits-per-sample decomposition; both degrade; near-zero-pass-rate regime; soft-label distillation as a bits-per-sample multiplier.
• Created automated-ai-research-llm-capability-boundary (concept, medium): Jang's first-person split — agents strong at running/optimizing experiments, weak at next-experiment selection and dead-end escape; "if you can't evaluate it, you can't auto research it." Sober May-2026 counterpart to autoresearch-recursive-self-improvement (tension noted, framed chronologically — same boundary, Jang more sober on takeoff).
• Created mechanism mcts-per-move-target-to-llm-rl-inefficiency (4 steps, 6 citations; status active, medium-high conviction): canonical home for the MCTS-vs-LLM-RL chain.
• Created can-llms-choose-the-right-research-question (question, status open): the intelligence-explosion-timing crux — can agents choose the right question, not just run experiments.
• Created eric-jang (person, 10 quotes attributed across mcts-vs-llm-rl-credit-assignment, rl-information-inefficiency, automated-ai-research-llm-capability-boundary, can-llms-choose-the-right-research-question, and the mechanism): AI researcher (robotics/RL), ex-1X VP of AI, ex-DeepMind Robotics. Aliased "Eric Zhang" (Dwarkesh's spoken-intro transcription slip).
• Updated dwarkesh-patel (cross-context — physical home: vault/projects/stock-market/wiki/entities/, tag thread/artificial-intelligence already present, +3 quotes on rl-information-inefficiency and mcts-vs-llm-rl-credit-assignment — the bits-per-flop framing and the "supervision through a straw" LLM contrast).
• Updated index: added eric-jang under People; new "RL and training methodology" Concepts subsection + new "Mechanisms" section; added can-llms-choose-the-right-research-question under Open Questions (first standalone question page in the thread); added the source under Sources.

Cross-links: the three concepts and the mechanism cross-reference andrej-karpathy (origin of "supervision through a straw") and autoresearch-recursive-self-improvement (Karpathy's parallel March-2026 AutoResearch account).

Sponsor handling: transcript lines ~271-283 (Jane Street data-center tour) carry mislabeled diarization (sponsor read attributed across both speakers) — treated as non-claim, no attribution drawn from it.

Contradictions surfaced: 0 hard contradictions. One tracked tension (recorded in automated-ai-research-llm-capability-boundary, not silently reconciled): Jang's May-2026 "direction-selection and dead-end escape still fail, and compute-multipliers don't stack" is more sober on near-term takeoff than Karpathy's March-2026 AutoResearch bullishness (autoresearch-recursive-self-improvement). Both locate the frontier at "verifiable inner loop yes, research taste no," so framed chronologically rather than as a coeval contradiction per the thread's vintage discipline.

Multi-context source (vault/sources/) — also tagged for projects/stock-market and projects/career. Speaker-aware ingest with named speaker Nick Turley (Head of ChatGPT at OpenAI) via AssemblyAI Speaker Identification.

Speaker-mapping correction. Same correction as the prior Bg2 Pod ingest: speaker B is mapped to "Brad Gerstner" in the frontmatter but is actually Apoorv Agrawal per show notes. Wiki attributes those quotes to apoorv-agrawal, not brad-gerstner.

Significance. This ingest adds the fourth independent practitioner vantage on the late-2025 / early-2026 coding-agent capability inflection: Karpathy (frontier-research, December) → Andreessen (VC, May) → Benioff (enterprise CEO, May) → Turley (AI-lab product head, March). The four vantages collectively make the inflection one of the highest-confidence findings in the AI thread. Turley is also the wiki's primary OpenAI-product-side voice — closes a meaningful gap relative to the wiki's prior Anthropic-heaviness.

• Updated apoorv-agrawal (cross-context within this thread — same physical home; +1 quote on OpenAI Code Red Q4 2025 context).
• Created nick-turley (person, 8 quotes attributed across ai-coding-productivity-paradox, ai-coding-agent-asymmetry-on-novel-code, chatgpt-super-assistant-vision, agi-timeline-decade-of-agents, gpu-as-zero-sum-constraint): Head of ChatGPT at OpenAI, ~3.5yr tenure; came from Instacart product leadership. Highest-ranking ChatGPT product voice in the wiki.
• Created openai (entity, subtype: company): the wiki's first dedicated OpenAI page. Anchored on Turley's framing of 900M WAU scale, super-assistant product vision, Q4 2025 Code Red, pricing-evolution forecast, and the competitive positioning vs Anthropic/Gemini. Notes the May 2026 Apple integration dispute (cross-link to existing anthropic coverage) and the broader capital-structure questions Chamath's SPV-crackdown framing raises.
• Created chatgpt-super-assistant-vision (concept, status: active, conviction: medium): the two-compounding-capabilities framing (actions × proactivity) + the escape-velocity-threshold model + why-code-first explanation (RL-friendly testability). The wiki's clearest operational characterization of what "super assistant" means.
• Updated ai-coding-productivity-paradox (concept): added "Fourth-vantage corroboration — the OpenAI product side (March 2026)" subsection with Turley's "engineers who don't open their IDE like ever" datapoint and his RL-friendliness-as-mechanism framing. Also folded in his consumer-Codex AGI-feel quote (mass-market analog to Karpathy's frontier-practitioner December inflection).
• Updated agi-timeline-decade-of-agents (concept): added "Turley AGI-feels (Mar 15, 2026) — OpenAI-product-side framing" subsection within the new N-point snapshot structure. Includes the GPT-4 → GPT-4o cost-curve datapoint, the reasoning-model-emergent-swearing AGI-moment, and the consumer-Codex transformation observation.
• Updated anthropic (entity): added "OpenAI Code Red context (Q4 2025) — Benioff's temporary Gemini swing" subsection — chronological framing for the Anthropic competitive trajectory. Benioff was vocally Gemini in Q4 2025; by May 2026 he was at $300M/yr Anthropic spend. The post-December-2025 coding-agent capability inflection landed on Anthropic's bet and re-anchored Benioff within ~5 months. The speed of the re-anchoring is a useful prior for the enterprise-buyer loyalty dynamics in this market.
• Updated index: added nick-turley under People; added openai under Labs; added chatgpt-super-assistant-vision under "Timelines and macro-AI framings" Concepts; added new "Operational mechanics" Concepts subsection referencing gpu-as-zero-sum-constraint and ai-as-services-disruption (both currently treated within parent pages, not yet standalone — flagged for future split-out if more evidence accrues); added the source under Sources with explicit "multi-context" annotation and the speaker-mapping correction noted.

Cross-context handoff: also pending in projects/stock-market (OpenAI commercial trajectory, GPU-zero-sum constraint as competitive dynamic, ChatGPT-shopping-as-discovery-channel, the Apple integration dispute for AAPL/MSFT positioning) and projects/career (Turley's "curiosity is the perma-skill" + writing-remains-important framing for career-runway thinking; the consumer-Codex "people who never coded make stuff" datapoint as an indicator of how fast vibe-coding tier-4 tools will accrete adoption).

Contradictions surfaced: 0. Turley's framings are entirely additive to the existing thread evidence — no claims that conflict with the wiki's prior view of the capability transition, the GPU constraint, or the competitive landscape.

Multi-context source (vault/sources/) — also tagged for projects/career. Speaker-aware ingest with named speakers Ali Ghodsi (Databricks CEO) + Arvind Jain (Glean CEO) via AssemblyAI Speaker Identification.

Speaker-mapping correction (important). The source's frontmatter speaker_mapping has C → Brad Gerstner because Brad was passed as the host hint for the Bg2 Pod, but the actual interviewer on this episode is Apoorv Agrawal (Altimeter partner, per show notes). Wiki attribution uses the corrected name — speaker C's quotes are attributed to apoorv-agrawal, not brad-gerstner. The source frontmatter stays as-is (immutable post-promotion).

Significance. This ingest fills the chronological middle (Dec 23, 2025) between Karpathy's Oct 17, 2025 "Decade of Agents" snapshot and Karpathy's Mar 20, 2026 "December inflection" update. Two enterprise-AI vendor CEOs (Ghodsi + Jain) give the deploy-side view at the date Karpathy later named as the personal-workflow inflection — and they don't yet name the inflection. Ghodsi explicitly says coding is "a little bit overhyped" Dec 23, 2025. Useful as a calibration point for inflection-propagation latency from frontier-tools side to general enterprise-AI deploy-side framings.

• Created ali-ghodsi (person, 5 quotes attributed across agi-timeline-decade-of-agents, llm-as-commodity-thesis, three-camps-of-ai, ai-coding-productivity-paradox): Databricks CEO + co-founder; UC Berkeley AMPLab alum. New person entity tagged for thread/artificial-intelligence. The wiki's primary anchor for Camp-3 / value-capture framings.
• Created arvind-jain (person, 4 quotes attributed across llm-as-commodity-thesis, ai-as-services-disruption, chatgpt-super-assistant-vision): Glean CEO + co-founder; former Distinguished Engineer at Google + Rubrik co-founder. The proactive-AI "personal companion" product-vision voice.
• Created apoorv-agrawal (person, 2 quotes attributed in this source; 1 more from the next ingest below): Altimeter partner; Bg2 Pod guest interviewer. With explicit speaker-mapping-correction note in the entity summary.
• Created databricks (entity, subtype: company): Enterprise data platform + AI deployment vendor; ~10,000-person org.
• Created glean (entity, subtype: company): Enterprise AI search / "personal companion" platform; $200M ARR by Q4 2025; failed-fine-tuning evidence supporting the commodity thesis.
• Created llm-as-commodity-thesis (concept, status: active, conviction: medium-high): Ghodsi's framing — frontier LLMs are commodities at the unit level; durable value accrues to proprietary data + workflow integration + the app layer. Glean's failed fine-tuning is direct first-person operator evidence.
• Created three-camps-of-ai (concept, status: active, conviction: medium): Ghodsi's superintelligence-quest / sober-researchers / value-capture partition. Useful as a map for sorting which camp's success criterion applies to a given AI debate (timelines, capex, AGI definitions, regulation).
• Updated agi-timeline-decade-of-agents (concept): added "December 2025 mid-period datapoints — Camp 3 operator view" subsection with Ghodsi's "we have AGI" framing, his continual-learning bottleneck corroboration of Karpathy's October framing, and his pre-inflection "coding is overhyped" view. Bumped the vintage banner from "Two-point snapshot (Oct + Mar)" to "N-point snapshot (Oct + Dec + Mar)".
• Updated ai-coding-productivity-paradox (concept): added "Pre-inflection sober deploy-side framing (Dec 2025)" subsection with Ghodsi's "coding is a little bit overhyped" datapoint as chronological calibration on how long it takes the inflection to propagate from frontier-tools to enterprise-AI-vendor framings.
• Updated anthropic (entity): added "Enterprise-vendor sentiment baseline (December 2025)" subsection with Ghodsi/Jain's rapid-fire "Anthropic stock + revenue up next 12 months" forecast. Useful as a Dec 2025 baseline against which the May 2026 $44B-ARR trajectory framing emerged.
• Updated index: added ali-ghodsi, arvind-jain, apoorv-agrawal under People; added new "Enterprise-AI vendors" subsection with databricks, glean; added llm-as-commodity-thesis and three-camps-of-ai under Concepts; added the source under Sources with explicit "multi-context" annotation and the speaker-mapping correction noted.

Cross-context handoff: also pending in projects/career. The career ingest will lean on the services-industry-disruption framing (relevant to the FE-career-runway question — services-industry capture is where the AI revenue lives, not software-TAM extension) and Arvind's first-person operator-routine evidence (Daily-Prep agent, change of instinct on when to consult AI vs. people).

Contradictions surfaced: 0 (Ghodsi's "we have AGI" view is not a contradiction with Karpathy's "Decade of Agents" — they're talking past each other on AGI definitions, which is exactly the kind of cross-camp confusion three-camps-of-ai is designed to clarify).

Multi-context source (vault/sources/) — also tagged for projects/career, threads/politics, projects/stock-market. Speaker-aware ingest with named speakers Jason Calacanis, Chamath Palihapitiya, Marc Benioff, David Friedberg via AssemblyAI Speaker Identification. (David Sacks + Brad Gerstner supplied as hints but not on this episode — extractor correctly didn't force their names onto unmatched voices.)

Significance for this thread: Three independent practitioners now converge on naming a Q4 2025 / early 2026 AI-coding-agent capability inflection (Karpathy December 2025; now Benioff "Anthropic 4.6 hit, boom"; plus Andreessen's earlier broader "20× productivity" framing). The convergence across practitioner-side (Karpathy), deploy-side enterprise CEO (Benioff), and VC-side (Andreessen, Sarah Guo) is high-signal triangulation that the inflection is structural rather than idiosyncratic.

• Created marc-benioff (person, 5 quotes attributed across anthropic, saaspocalypse-thesis, ai-coding-productivity-paradox): Salesforce CEO + frontier-AI buy-side practitioner. Physical home is the AI thread per his frontier-AI-practitioner status; will cross-context-tag for career and stock-market on those ingest passes.
• Cross-context tag-append: chamath-palihapitiya (physical home vault/projects/stock-market/wiki/entities/, +1 quote on Anthropic-SPV-crackdown framing). david-friedberg (same physical home, +1 quote on assistant-interface competitive landscape).
• Created saaspocalypse-thesis (concept, status: active, conviction: medium): captures the AI-eats-SaaS bear case ($180B combined market cap lost, 2× sales multiples) and Benioff's "Not My First SaaSpocalypse" operator-side rebuttal. Substantial enterprise-deploy-side complement to Karpathy's practitioner-side framing. Cross-linked to ai-vampire-pattern and ai-macro-signals-2026 in the career thread.
• Updated anthropic (entity): substantial additions — new "Enterprise deploy-side adoption (May 2026)" subsection with Benioff's $300M/yr Salesforce spend datapoint, the "rocket ship" framing, the "Anthropic 4.6 hit, boom" inflection corroborating Karpathy's December 2025 framing, and Benioff's competitive-positioning narrative of why Anthropic won the coding-agent layer. New "OpenAI vs Apple ChatGPT integration dispute (May 2026)" subsection with Apple-as-potential-Anthropic-customer speculation per Friedberg. New "Capital structure / SPV crackdown (May 2026)" subsection with Chamath's framing.
• Updated ai-coding-productivity-paradox (concept): added "May 2026 corroboration from a SaaS-buyer perspective (Benioff)" subsection. Three different practitioners (Karpathy / Andreessen / Benioff) from three different vantage points (frontier research / VC / enterprise CEO) all naming the same Q4 2025 / early 2026 inflection is now well-documented.
• Updated index: added marc-benioff under People; added cross-context chamath-palihapitiya and david-friedberg under People (the AI thread now has 4 cross-context People entries from stock-market — the All-In hosts plus Dwarkesh from earlier); added saaspocalypse-thesis under "Timelines and macro-AI framings" Concepts; added source entry under Sources.

Cross-context handoff: this is the first of four target contexts. career / politics / stock-market are pending in this same auto-chain. Cross-thread implications already wired:

• saaspocalypse-thesis is the AI-thread enterprise complement to the career thread's ai-vampire-pattern and ai-macro-signals-2026 Theme 2 (structural decline in software employment) and Theme 7 (Karpathy operator-skeptic counter).
• The Benioff "Anthropic 4.6 hit, boom" framing corroborates the agi-timeline-decade-of-agents two-point snapshot's December 2025 inflection date.
• The OpenAI-vs-Apple dispute and Anthropic-SPV-crackdown subsections will be referenced by the stock-market ingest for the AAPL / MSFT / OpenAI-IPO / Anthropic-IPO mechanics.

Contradictions surfaced: 0 (Benioff's SaaSpocalypse rebuttal is a tension with the market-pricing bear case, but it's framed as both-true-pending-data-resolution on the concept page rather than a queue-blocking conflict).

Multi-context source (vault/sources/) — also tagged for projects/career. Speaker-aware ingest with named speakers Andrej Karpathy + Sarah Guo via AssemblyAI Speaker Identification.

Significance. This is the wiki's first capability-refresh ingest under the new ../../_meta/AI_CAPABILITY_TRACKING discipline. Karpathy was flagged in the meta doc as "refresh due" (last source Oct 2025). This source — recorded March 20, 2026, 5 months later — substantially updates his framing of current AI-coding-agent capability while leaving his timeline framing intact. The wiki updates here are explicitly chronological per the discipline.

• Updated andrej-karpathy (cross-context — physical home: threads/artificial-intelligence, already tagged for thread/artificial-intelligence, +12 quotes attributed). Karpathy entity now holds 33 quotes total across two recording dates 5 months apart: 21 from October 2025 (Dwarkesh) and 12 from March 2026 (No Priors).
• Created sarah-guo (person, 2 quotes attributed): Conviction VC founder + No Priors co-host. New person entity tagged for thread/artificial-intelligence.
• Created autoresearch-recursive-self-improvement (concept, status: active, conviction: medium): captures Karpathy's first-person demonstration of recursive-self-improvement on his own nanochat model — AutoResearch found hyperparameter tunings he missed after two decades of manual practice. Frames the surviving asymmetry (research-direction still human; verifiable-evaluation loops automatable) and the structural untrusted-swarm framing (cheap-to-verify, expensive-to-search → blockchain / Folding@home analog). Highest-leverage AI-2027 / recursive-self-improvement evidence in the wiki.
• Updated agi-timeline-decade-of-agents (concept): added a major "Two-point snapshot: Karpathy's own update Oct 2025 → Mar 2026" section explaining what stayed (timeline framing, jaggedness, march-of-nines) and what changed (December 2025 inflection, bottleneck reframed as orchestration-skill, asymmetry narrowed substantially, AutoResearch active, jobs framing flipped to Jevons-paradox optimism). Updated vintage banner to "two-point snapshot" format. Conviction stays at medium but with explicit note about the framing's capability-state component being superseded.
• Updated ai-coding-agent-asymmetry-on-novel-code (concept): conviction downgraded medium-high → low-medium. Added "March 2026 update — the asymmetry narrowed substantially" subsection in Evidence. The structural claim (verifiable vs. soft domains) survives; the operational claim (autocomplete is the productive mode for senior practitioners) doesn't.
• Updated ai-coding-productivity-paradox (concept): restructured the Karpathy subsection into explicit Oct 2025 baseline + Mar 2026 update format. Added Sarah Guo's independent corroboration of the multi-agent orchestration workflow shift from Conviction's portfolio.
• Updated index: added sarah-guo under People; added autoresearch-recursive-self-improvement under "Timelines and macro-AI framings" Concepts; updated existing concept entries with chronological framing notes; added the new March 2026 source under Sources with explicit cross-reference to the October source it updates.
• Updated ../../_meta/AI_CAPABILITY_TRACKING: Karpathy's refresh-due status moved from "Refresh due" → "Recent (refreshed 2026-05-15)", with an explicit note that this is the discipline's first applied case study.

Cross-context handoff: also pending in projects/career. The career ingest will lean on the Jevons-paradox framing, the December-2025 inflection as a runway-relevant signal, and the BLS-jobs-data analysis Karpathy mentions. Karpathy's ai-vampire-pattern "contradiction" with Andreessen (logged 2026-05-15 morning) now needs re-framing in light of this — Karpathy's March view substantially aligns with Andreessen's productivity-gain framing.

Contradictions surfaced: 0 (the October→March shift is not a contradiction with prior wiki claims; it's exactly the chronological-update pattern the new vintage discipline is designed to handle).

Multi-context source (vault/sources/) — also tagged for threads/autonomous-driving and projects/career. Speaker-aware mode: speakers Andrej Karpathy and Dwarkesh Patel, named via AssemblyAI Speaker Identification with host=Dwarkesh Patel + auto-extracted guest hint Andrej Karpathy from the title. 264 named turns; speaker_mapping {A: Andrej Karpathy, B: Dwarkesh Patel}.

• Created andrej-karpathy (person, 10 quotes attributed across agi-timeline-decade-of-agents, ai-coding-agent-asymmetry-on-novel-code, ai-coding-productivity-paradox, ai-coding-tool-landscape-2026). Roles: Eureka Labs founder, former Tesla AI director (Autopilot lead 2017-2022), early OpenAI co-founder, Hinton-lineage deep-learning practitioner.
• Updated dwarkesh-patel (cross-context — physical home: vault/projects/stock-market/wiki/entities/, tag thread/artificial-intelligence appended, +2 quotes attributed): the pre-training-vs-in-context-learning 35-million-fold information-bandwidth claim and the compiler-productivity-vs-programmer-automation framing.
• Created agi-timeline-decade-of-agents (concept, status: active, conviction: medium): captures Karpathy's central thesis — "Decade of Agents not Year of Agents", concrete cognitive deficits in current agents (continual learning, multimodality, computer use), "ghosts not animals" pre-training-vs-evolution distinction, RL as "sucking supervision through a straw" with LLM-judge gameability, march-of-nines demo-to-product gap, and "cognitive core" as the open research direction. Cross-thread tension flagged with anthropic's revenue trajectory.
• Created ai-coding-agent-asymmetry-on-novel-code (concept, status: active, conviction: medium-high): captures the first-person nanochat observation — agents win on boilerplate / familiar-pattern / unfamiliar-language scaffolding, lose on idiosyncratic novel-architecture work. Framed as the concrete mechanism that would block the AI-2027 / recursive-self-improvement loop (frontier-lab work is by definition off-data-manifold). Cross-linked to agi-timeline-decade-of-agents and ai-coding-productivity-paradox.
• Updated ai-coding-productivity-paradox (concept): added a "First-person operator framing" subsection in Evidence with Karpathy's three-tier developer-interaction model and the boilerplate-vs-novel asymmetry as one possible reconciliation of the METR (−19% slowdown) vs Uvik (2-3× speedup) tension.
• Updated anthropic (entity): added "Practitioner adoption signal (October 2025)" subsection with Karpathy's "I use Daily Claude and Codex" datapoint — small but real evidence on the agentic-coding-tool layer alongside the existing commercial-trajectory evidence.
• Updated index: added new "Timelines and macro-AI framings" subsection under Concepts; added andrej-karpathy and cross-context dwarkesh-patel to People; added source entry under Sources.

Cross-context handoff: this source is also pending in threads/autonomous-driving and projects/career; each will record its own ingest separately. The autonomous-driving ingest will lean on Karpathy's first-person Tesla Autopilot framing (5 years for 2-3 nines; "self-driving is not even near done"; Waymo capex unprofitability and teleoperation in the loop). The career ingest will pick up the macro-AI timeline framing (Decade of Agents) and the developer-productivity-tier model for the FE-career track.

Contradictions surfaced: 0 (cross-thread tension with anthropic revenue trajectory is logged on the concept page as both-true asymmetric, not as a contradiction).

Multi-context source (vault/sources/) — also tagged for projects/stock-market and projects/career. AssemblyAI A/B/C/D unnamed heuristic-fallback labels — source-attributed mode (no person-entity creation). All-In E272 (Calacanis + Sax + Chamath + Brad).

• Created anthropic (entity, subtype: company): frontier-lab entity page covering commercial trajectory ($10B→$30B→$44B ARR Q1-April 2026 with Brad's compute-constraint caveat against $1T-by-2027), compute supply (SpaceX-leased Colossus 1: 220K GPUs + 300MW), safety/regulation posture (Mythos cyber-capability triggered FDA-for-AI debate; self-regulation vs. regulatory-capture both-readings), and competitive framing (Sax's "Anthropic = porcupine, OpenAI = fox" frame on focus discipline).
• Updated index: added a new "Labs" section under Entities with anthropic (positioning the thread to widen beyond tools to frontier-lab dynamics per the SCOPE update from 2026-05-13). Added source entry under Sources.
• No new concept pages. The AI-regulation / FDA-for-AI debate is interesting and within scope but one source isn't enough to scaffold a concept page yet — track for future ingests on the same theme. Same logic for the cyber-capable-frontier-models dimension (Mythos vs. OpenAI's equivalent).
• Cross-context handoff: stock-market and career independently ingested the same source on 2026-05-14 with their own lens (stock-market focused on the EWS/Anthropic compute deal as cash-flow bridge for SpaceX IPO; career focused on the political-backlash dimension as Theme 6 update to ai-macro-signals-2026).
• Contradictions surfaced: 0.

Re-diarized after the AssemblyAI Speaker Identification API shape was fixed in .claude/skills/_lib/diarize.py (the old speaker_options.speakers parameter was silently ignored by the API; the current shape is speech_understanding.request.speaker_identification.{speaker_type: name, speakers: [{name, description}]}). All 4 speakers in this episode were correctly named: A=David Sacks, B=Brad Gerstner, C=Chamath Palihapitiya, D=Jason Calacanis. The source file at vault/sources/2026-05-08-all-in-podcast-elon-s-anthropic-deal-the-next-ai-monopoly.md has been rewritten with named-speaker turns and a rediarized_at frontmatter field; this log entry is the audit trail for this context's view.

No wiki-citation rewrites in this context — the earlier ingest used source-attribution (From [[source]]:) because A/B/C/D were heuristic-fallback labels, which doesn't break or improve with named speakers. A future re-ingest of this source (if/when motivated) would now produce speaker-attributed evidence ([[david-sacks]] in [[source]]:). Cost: $0.2324; budget now $3.14 / $25 for 2026-05.

Ingested @ClaudeDevs's X announcement of /ultrareview, a new Claude Code slash command (research preview) that runs a fleet of bug-hunting agents in the cloud and delivers findings back to the CLI or Desktop. Recommended for use before merging critical changes (auth, data migrations). Pro/Max users get 3 free reviews through 5/5 2026. Single-source ingest with rich reply-thread commentary — reception is polarized.

• Created multi-agent-code-review (concept): the underlying pattern (fleet-based pre-merge review) distinct from the specific product. Captures the "fleet catches interactions that single-pass review misses" thesis from @MindTheGapMTG, the strategic framing from @Surreal_Intel ("permanent machine review layer between 'done' and 'deployed'") and @Tahseen_Rahman ("review has to become parallelized and adversarial"), plus the Witness adjacent-tool data point. Multiple open questions: will the cultural-expectation shift @jatingargiitk predicts actually happen? Does fleet-review outperform single-agent review on real bug classes? What's the post-free-window pricing story?
• Updated claude-code: added a "Later-shipped features" section with the /ultrareview details, the full polarized reception from the reply thread (positive pattern-legitimizing takes, the cultural-norm-shift claim as open question, the @thepatriotvlls negative first-day report, the usage-limit theme connecting to the already-documented Pro-vs-heavy-use cost gap). Added source and cross-link to multi-agent-code-review.
• Updated index.md: +1 concept (multi-agent-code-review) under Claude Code-specific workflow, +1 source entry.
• Deferred / not created:
  • No entity page for Witness — single-comment mention from the tool's author, no independent corroboration. Captured as a data point in multi-agent-code-review.
  • No entity pages for reply commenters (Miles Deutscher precedent — commenters aren't entities).
  • No update to ai-coding-tool-landscape-2026 — light-touch decision. If fleet-review becomes a recurring theme across more sources (other vendors shipping similar features, independent benchmarks, etc.), revisit the landscape page then.
• No contradictions against existing wiki claims. The announcement reinforces the already-documented pricing/usage-limit pattern on claude-code; it doesn't contradict any prior claim.
• Note on source quality: reply-thread sample is self-selected and skewed toward strong opinions. Positive and negative takes alike are explicitly labeled as commenter opinions rather than endorsed wiki claims.

Ingested the deep-dive autoresearch on Lovable / Bolt / v0 / Replit Agent. Four tier-4 tools graduated from concept-page references to their own entity pages, plus one new concept page for the security picture.

• Created lovable (entity, tier-4): dominant tier-4 tool by app count (~4K of 5,600 in Escape.tech scan); corrected pricing ($25/mo Pro, not $39 as in first autoresearch); Visual edits are the only zero-credit affordance in the tier.
• Created bolt-new (entity, tier-4): StackBlitz WebContainers architecture; lowest stack lock-in; token-per-project-size economics; practitioner review documenting 2M+ token burn on a single dashboard build.
• Created v0-vercel (entity, tier-4): slug disambiguated from v0 to avoid any future basename fragility. Captures the Feb 3, 2026 relaunch moving v0 from "prototype toy" to production infrastructure (Git panel, GitHub repo import, database integrations, agentic workflows). Best code-quality ranking in the tier per one practitioner review.
• Created replit-agent (entity, tier-4): dedicated page for the tier's most autonomous and most cost-volatile tool. Full Agent 3 capability list (200-min runtime, self-testing, subagents, effort modes). Full coverage of the September 2025 cost-escalation crisis (The Register $1K/week user reports + root causes in effort-based pricing and subagent billing) and the July 2025 SaaStr / Jason Lemkin production-database deletion incident (AI admitted ignoring freeze directive; 1,206 executive records wiped).
• Created vibe-coding-security (concept): Escape.tech's systematic scan of 5,600+ vibe-coded apps → 2,000+ vulnerabilities, 400+ exposed secrets, 175 PII instances (medical records, IBANs). Dominant pattern: exposed Supabase JWT + misconfigured RLS. Adjacent vendor CVEs from Kaspersky: Cursor CVE-2025-54135, Anthropic MCP CVE-2025-53109, Windsurf prompt-injection, Amazon Q near-incident. Novel attack class documented: package hallucination → typosquatting.
• Updated vibe-coding-app-builders: replaced per-tool sub-sections with entity-page links + corrected pricing + Security section + Cost-volatility section. Added update banner at top explaining the graduation.
• Updated ai-coding-tool-landscape-2026: Tier 4 entries now link to the new entity pages with concise corrected descriptions.
• Contradictions preserved (not reconciled):
  • Lovable Pro pricing: first autoresearch said $39/mo; deep-dive confirmed $25/mo from Lovable's primary pricing page. Wiki now uses the primary-source $25/mo figure; noted the correction in vibe-coding-app-builders.
  • Escape.tech sample is 70%+ Lovable because Lovable has more public apps, not because Lovable is less secure than peers. Surfaced as an open question in vibe-coding-security.
• Deferred: Factory (still no source), Aider, Cline, Amazon Q, Augment, JetBrains Junie, AWS Kiro — referenced but not profiled.

Ingested Miles Deutscher's promotional thread on Anthropic's new Claude Design platform. Thin on verifiable product detail but rich in polarized user reactions.

• Created claude-design (entity): Anthropic's April 2026 AI design platform. Captured the "quiet launch, strong product" framing from the original tweet; the overwhelming reply-thread signal about usage-limit frustration (dominant complaint); the quality critiques ("just pre-made stuff", "SVG support terrible", "QA is nil", "descending cycles it can't recover from"); and the one positive data point from @OneManSaas about component variations saving hours against Figma.
• Surfaced but did not resolve: pricing/credit tiers are not stated in the source. Users uniformly complaining about hitting limits in 4 prompts / 2 designs / a few slides. Flagged as the primary open question on the entity.
• Did NOT create entity pages for Miles Deutscher, OneManSaas, or other reply commenters — they're source authors/commenters, not entities with durable reference value. Source citation captures attribution.
• Note on source quality: The Miles Deutscher thread is influencer-style promotional content (cover-image teaser, "ultimate guide" framing) — the positive framing in the original post is weighed against the detailed critical replies in the thread. This is flagged explicitly in the claude-design weaknesses section.

Ingested Meng To's short thread on neuform.ai — a free template library designed for use with Claude Design.

• Created neuform-ai (entity): 400+ free design-system templates as HTML/DESIGN.md; primary use case is feeding claude-design with high-quality template context. Includes webgl/threejs landing-page sections.
• Noted the filename-vs-product-name typo: the clipping was named neuroform-design-systems but the product is neuform.ai (no "r"). Wiki entity page and cross-links use neuform-ai; source citation preserves the filename.
• Cross-linked neuform-ai ↔ claude-design. Meng To's workflow — "baseline AI output is generic, feed it a high-quality template instead" — offers a counter to the claude-design "it's just pre-made stuff" critique from a different angle: the generic baseline is the problem, external templates are the workaround.
• Did NOT create a Meng To entity page — author, not entity. Durable identity lives in the [[@MengTo]] frontmatter and in the source basename.
• Single-source-evidence page; flagged multiple open questions around template quality, adoption, licensing, sustainability.

Ingested the /autoresearch "best ai coding tools" synthesis — a three-round, 14-source web pass covering the whole April 2026 AI coding tool market. Substantial expansion of the wiki: 7 new entity pages, 5 new concept pages, 3 existing pages updated with new market-scale and benchmark data.

• Created cursor (entity, tier-1 IDE): $1B ARR claim, 360K paying users, 18% JetBrains adoption with stalled growth, pricing, METR RCT's test stack.
• Created github-copilot (entity, tier-1 IDE): Microsoft's broad-distribution tool; 29% adoption / 76% awareness / 15M devs / Fortune 100 90%. Noted Copilot Coding Agent as an in-family tier-3 sibling without its own page yet.
• Created windsurf (entity, tier-1 IDE): distant third; March 2026 price raise to match Cursor; 8% adoption. Thin-evidence page by design.
• Created codex-cli (entity, tier-2 terminal): OpenAI's throughput-leader; Terminal-Bench top; bundled into ChatGPT; 3% surveyed adoption likely undercounted.
• Created devin (entity, tier-3 background agent): Cognition's canonical autonomous agent; 67% PR merge rate — the rare output-oriented metric in the source.
• Created google-antigravity (entity, tier-3): November 2025 agent-first platform; dual Editor/Manager view; Gemini 3 Pro default with BYO Claude/GPT-OSS; free in public preview; March 2026 AgentKit 2.0 ships 16 specialized agents.
• Created google-jules (entity, tier-3): Google's GitHub-native background agent; 2026 updates added proactive/scheduled/Render features; Jitro V2 rumor flagged as weak evidence.
• Created ai-coding-tool-landscape-2026 (concept): the four-tier framing — central organizing idea for this thread's market view. Includes the Cursor-vs-Claude-Code ARR contradiction, the fuzzy background-agent boundary, and the cross-survey adoption disagreement.
• Created ai-coding-tool-stacking (concept): the 70%-use-2-to-4-tools pattern; Uvik's 2–3× time-to-merged-PR field data; Uvik's 1.5–2× bug rate on legacy work caveat; relationship to Boris's intra-tool parallelism.
• Created ai-coding-productivity-paradox (concept): full arc of METR's July 2025 "19% slower" RCT → February 2026 walk-back → unresolved state; MIT Tech Review's skeptical voices; Satya-vs-empirical tension. Deliberately heavy on contradictions; flagged as the contested-evidence anchor for future ingests.
• Created ai-coding-benchmarks (concept): SWE-bench Verified/Pro/SEAL, Terminal-Bench, Aider Polyglot, LiveCodeBench, HumanEval/MBPP with contamination/saturation notes. Captures MorphLLM's "same model scores differently in different agents" finding — the scaffolding-vs-model point.
• Created vibe-coding-app-builders (concept): the tier-4 category (Lovable, Bolt.new, v0, Replit Agent, plus Bubble/Softr/Base44/Adalo/Dyad in brief). Did NOT create separate entity pages for the vibe-coding tools — they're listed inside this concept and graduate to entities only if a second source substantiates them.
• Updated claude-code: added market-position section ($2.5B ARR per Uvik, 18% adoption with 6× growth, 60–70% of teams via bottom-up adoption), pricing detail ($150–200 real heavy-use vs. $20 advertised), benchmark scores, and a note that the METR RCT used Cursor+Sonnet, not Claude Code.
• Updated claude-opus-4-5: added benchmark-position section (~80.9% SWE-bench Verified, ~45.9% SWE-bench Pro SEAL, ~85% Aider Polyglot for successor 4.6), plus the OpenAI-confirmed contamination caveat on Verified. Also noted 4.6/4.7/Mythos Preview exist and outscore 4.5 — 4.5 remains important only because Boris's workflow argument is built around it.
• Updated parallel-claude-workflow: added "how this relates to multi-tool stacking" section distinguishing Boris's intra-tool parallelism from the wider market's cross-tier stacking pattern; linked to ai-coding-tool-stacking and ai-coding-tool-landscape-2026.
• Contradictions surfaced (not silently reconciled — preserved in the relevant concept pages for future adjudication):
  • Cursor $1B ARR vs. Claude Code $2.5B ARR — both from round-up blogs, neither from primary disclosures. Flagged in ai-coding-tool-landscape-2026.
  • METR 2025 RCT (−19%) vs. Uvik 2026 field report (+200–300% on time-to-first-merged-PR). Populations, tools, and tasks differ; reconciliation is non-trivial. Captured in ai-coding-productivity-paradox and ai-coding-tool-stacking.
  • Single-tool adoption across surveys — JetBrains (18% each for Cursor and Claude Code) vs. Claude5.ai (28% Claude Code / 24% Cursor as primary-tool selections). Likely a primary-vs-any-use phrasing difference. Flagged in ai-coding-tool-landscape-2026.
  • Background-agent definition — Builder.io's five-point definition is clean but other reviews put Cursor and Claude Code inside the tier. Boundary is genuinely unstable.
• Deferred:
  • Lovable, Bolt.new, v0, Replit Agent not created as separate entity pages — mentioned within vibe-coding-app-builders. Graduate on second source.
  • Aider not created as an entity page — one-paragraph evidence in the source, mentioned in the landscape concept. Graduate on second source.
  • Factory named in round-2 search but no fetched source profiled it — flagged as gap in the autoresearch, not created here.
  • GitHub Copilot Coding Agent mentioned inside github-copilot and ai-coding-tool-landscape-2026 but not as its own entity page.
• Non-obvious vault-wide check: Confirmed no basename collisions between these 12 new pages and anything in projects/career/wiki/ (which has its own set of AI-related pages: ai-implementer-opportunity, ai-macro-signals-2026, ai-native-multi-agent-workflow, etc.). The career pages are career-scoped and distinct; no cross-context link created in this ingest.

Ingested first source — @DataChaz's 11-minute interview with Boris Cherny (creator of Claude Code) on his daily Claude Code workflow (15 tips). Transcript of the embedded video was folded in via transcribe-clipping before promotion. Seeded the thread wiki with three entity pages and four concept pages.

• Created boris-cherny (entity): person, creator of Claude Code; dogfooding positions and notable-quote index.
• Created claude-code (entity): the product surface itself (terminal, web, iOS, Android; plan mode, auto-accept edits, CLAUDE.md, GitHub Action); noted the one-directional iOS ↔ desktop sync limitation flagged by a reply commenter.
• Created claude-opus-4-5 (entity): the model Boris uses for everything; flagged his "smarter model is cheaper end-to-end" claim as an open question because the source has no token-count numbers.
• Created plan-then-execute-coding (concept): "Once the plan is good, the code is good." Coupled to parallel-claude-workflow via Boris's tab-cycling routine.
• Created parallel-claude-workflow (concept): 5–10 Claudes concurrently; separate checkouts, no worktrees; mobile as a first-class coding surface. Cross-corroborated by reply commenters @MGMurray1 and @PangestuAden.
• Created claude-md-team-knowledge-base (concept): single CLAUDE.md per repo, multiple edits per week; Boris's "compound engineering" framing credited to Dan Shipper; noted @MGMurray1's 7-file variant (SOUL.md, AGENTS.md, …) as a contrasting data point.
• Created agent-output-verification (concept): tip #13 / one of Boris's top three performance levers; examples are all engineering-flavored (tests, server, Chrome extension) — flagged as an open question whether/how this translates to non-coding agent work.
• No contradictions surfaced within the source.
• Note: whisper-base transcript mis-hears "Claude" as "cloud"/"quad"/"clawed" and "spec-driven" as "spectraven"; quoted excerpts in wiki pages were normalized to the obvious intended terms. The verbatim transcript remains in the source file for provenance.
• Note: source body refers to "Dan Chipper's compounding engineering"; Boris corrects in the video that the concept is "compound engineering," and the person is Dan Shipper. Wiki pages use the corrected name.

Thread artificial-intelligence scaffolded via /thread-create.

• Scope: undefined — set in SCOPE.md when ready