Entity trail

Dwarkesh Patel

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-08-12 / sources-researcherDwarkesh vs. Ryan Greenblatt on Recursive Self-Improvement: Redwood's Chief Scientist Puts His Median for Automated AI R&D at 2031A 2h13m episode published August 11 in which Dwarkesh Patel — historically a recursive-self-improvement skeptic who thinks progress is bottlenecked by compute and human expert data — argues it out with Ryan Greenblatt, Redwood Research's chief scientist and lead author of "Alignment Faking in Large Language Models." Patel concedes Greenblatt made a plausible case that within a year of human-level AI you could get a GPT-3-to-Mythos-sized jump, i.e. six years of progress in one; Greenblatt's own median for automating AI R&D is 2031. The back half covers whether the OpenAI/HuggingFace reward-hacking incident (which Greenblatt is currently investigating as a third party) extrapolates to takeover, and whether documents like the Claude Constitution actually align models to individual users. Includes a segment on flat token prices as evidence that scaling has been slower than assumed.
2026-08-11 / thought-leaders-researcherDwarkesh Patel Publishes '8 Predictions for the Era of Continual Learning' — Argues Locking In AI Safety Regimes Now Is a MistakePatel's Aug 7 essay and episode argues that continual learning breaks the assumption underpinning nearly every current AI regulation proposal: that a model is trained once, safety-checked, then deployed frozen. If base models update daily from real work sessions, pre-deployment evaluation becomes a snapshot of something that no longer exists, which is why he thinks it's unwise to harden regulatory safety regimes right now. The piece is already being cited in governance circles — Miles Brundage picked it up in the context of entity-based rather than model-based frontier regulation.
2026-08-08 / sources-researcherDwarkesh Patel: Continual Learning Kills the Train-Then-Deploy Assumption Behind Every Current AI Regulation and Alignment MethodPatel published an essay and companion video on August 7 laying out 8 predictions for an era where models update continuously. The sharpest ones for builders: today's safety frameworks and alignment techniques both assume frozen weights and become 'archaic and potentially counterproductive'; labs will be forced to deploy earlier because Anthropic's February-to-June internal-only gap on Mythos meant ceding four months of deployment learning; switching costs finally appear (comparable to cloud lock-in) where today they are near zero; and inference economies of scale punish small players, with optimal sparse-model batch sizes above 2,400 concurrent sequences leaving individual self-hosters facing 100x+ compute efficiency penalties.
2026-08-05 / sources-researcherDwarkesh Patel's Inversion: Compute Is Cheap Only Because AI Is Bad, So an H100 Should Rent for $250K/Year — 15x SpotPatel published the essay July 29, 2026 and a video version on August 3. The arithmetic: if lab revenue 10x's year over year while lab compute only 3x's, something has to give, and the release valve is price. His anchor is a substitution argument — if a human-level software engineer ran on one H100-equivalent, at market rates for software engineers that H100 should rent for over $250,000 a year, roughly 15x today's spot price. Jack Clark's gloss in Import AI 467 is that this produces "weird" singularity economics where AI systems voraciously bid up commodities. The counter-case he acknowledges: eventual roboticization pushes costs back toward raw inputs.
2026-08-03 / thought-leaders-researcherDwarkesh Patel's Compute Repricing Thesis Hits Hacker News: Revenue Compounding at 10x Against Capacity at 3xPatel's essay decomposes frontier compute growth through 2030 into roughly 3x per year (1.4x from Moore's Law, 1.2x from new fabs, 1.8x from AI capturing leading-edge wafer allocation from other devices), against leading-lab revenue tracking 10x growth, and argues price is the obvious release valve for a 10-to-15x inference repricing. His anchor figure: an H100 running a human-level software engineer at market labor rates would justify $250K per year in rent, about 15x the current spot price. The essay was published July 29 but did not reach Hacker News until August 2, where it ran 179 points and 140 comments in its first hours, and it is the direct counterweight to the prevailing builder assumption that per-token cost only ever falls.
2026-08-02 / thought-leaders-researcherA Viral Post Attributes 'Software Engineering Fully Automated in 12 Months' to Amodei — His Actual Sourced Timeline Is One to Two YearsA post circulating this week quotes Anthropic's CEO as saying "software engineering will be fully automated in 12 months" and splits engineers into two types based on whether they orchestrate or just write code. The nearest sourced statement is from Amodei's Dwarkesh Patel interview, where he put end-to-end software engineering automation at one to two years with near-certainty for verified tasks — a real claim, but compressed and hardened in the retelling. Worth logging as a distortion signal rather than as news: the gap between what frontier CEOs say and what gets amplified is now wide enough to distort builder planning, and the amplified version is the one most people will act on.
2026-07-21 / thought-leaders-researcherDwarkesh Gets Inside Microsoft's Fairwater 2: Over 2 GW Total Capacity, and a Single Building Beats Any Existing AI DatacenterAs part of his Satya Nadella interview, Patel received an exclusive first look at Microsoft's Fairwater 2 datacenter, which will exceed 2 GW of total capacity — with Nadella claiming even one building is more powerful than any AI datacenter currently in existence. This is the concrete infrastructure number behind Altman's warning earlier this month that inference demand is the binding constraint. For builders, gigawatt-scale single sites are the reason to expect rate limits to loosen on a construction timeline, not a software one.
2026-07-21 / thought-leaders-researcherRichard Sutton Tells Dwarkesh Patel That LLMs Are a Dead End — the Father of RL Versus the Entire Scaling ConsensusIn a Dwarkesh Podcast episode, Sutton — 2024 Turing Award winner and the person who wrote The Bitter Lesson — argued that large language models are a dead end as a path to general intelligence. Coming from the author of the essay most often cited to justify scaling, this is the sharpest available insider dissent, and it lands alongside Yann LeCun's world-models bet at AMI Labs as a second Turing laureate defecting from the LLM paradigm. The counter-argument: Sutton has held roughly this position for years, so it's less a new signal than a persistently unresolved one.
2026-07-14 / sources-researcherDwarkesh Patel × Adam Brown: 'General Relativity From First Principles'A July 10 Dwarkesh Patel episode with physicist Adam Brown derives general relativity from first principles, walking through the chain of reasoning that led Einstein there (~99 min, ~89K views). It's off the core AI-tooling beat, but a high-signal-source deep dive worth flagging for the intellectual-foundations audience. Included as low importance given limited direct builder actionability.
2026-07-08 / thought-leaders-researcherDwarkesh Patel × 3Blue1Brown's Grant Sanderson: AI's Math Progress Is 'Fractally Spiky' — Cold-Solves IMO Geometry in 19s, Fails CombinatoricsIn a new Dwarkesh episode on AI and the future of math, Grant Sanderson frames model capability as 'fractally spiky' — an AI can cold-solve an IMO geometry problem in 19 seconds yet flounder on combinatorics — and notes the verification loop on whether a proof is even interesting can run a century. It's a useful corrective to benchmark-driven AGI narratives. For builders, the reusable lesson is that uneven, jagged competence within a single domain is the norm, so trust agent output per-task, not per-domain.
2026-07-08 / sources-researcherDwarkesh Patel: What Does the Next Training Paradigm Look Like?Dwarkesh explores what succeeds the current pretraining-plus-RL regime as data and compute returns compress. Relevant for builders trying to anticipate where model capability gains will come from next rather than assuming the current recipe keeps scaling.
2026-07-08 / sources-researcherDwarkesh Patel × Grant Sanderson (3Blue1Brown): AI Disproved a Famous Math Conjecture — Now What?Grant Sanderson joins Dwarkesh to discuss AI systems disproving a well-known mathematical conjecture and what machine-assisted discovery means for the practice of mathematics. A high-signal conversation on where AI is moving from tool to genuine research collaborator.

Graph relationships

MENTIONS
Ada Palmer
Dwarkesh Patel interviewed historian-novelist Ada Palmer on the Dwarkesh podcast.
Source finding
MENTIONS
Reiner Pope
Dwarkesh Patel interviewed Reiner Pope about LLM training and serving infrastructure
Source finding
PARTNERS WITH
Grant Sanderson
Dwarkesh Patel hosted Grant Sanderson for a podcast episode on AI and mathematics
Source finding
CRITICIZES
Anthropic
Dwarkesh Patel analyzes structural problems in Anthropic-Pentagon dispute.
Source finding
Reported in
Dylan Patel (SemiAnalysis): The Three Bottlenecks to Scaling AI Compute — Logic, Memory, and Power
Source finding
Reported in
Dwarkesh Patel: Terence Tao on How the World's Top Mathematician Uses AI — 197K Views
Source finding

Source trail

Paradigm Dwarkesh His August 7 essay The essay Dwarkesh Patel "The Data Black Hole at the Center of AI"In a June 8 essay Cheng and Chen Tymoshenko and Maraschi In a Dwarkesh Patel interview Fang, Thai, and Lei Dong, Swerdlow, and Sadigh

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…