Synthos Think Pieces · frontier science & second-order effects · July 11, 2026

AI is starting to build AI — the takeoff debate, and who's exposed

Recursive self-improvement is the idea that AI systems can now write the code, run the research and design the chips that make the next AI — a loop that, if it holds, compounds. The receipts are real but partial: Google's AlphaEvolve delivered a 23% speedup on part of Gemini's training stack, ~25% of the characters checked into Google's codebase are AI-generated (with human oversight), and coding agents draft-and-commit pull requests today. But the experts we track split hard — Jordi Visser says "we've hit recursive self-improvement… models now build models"; All-In and a gradual-diffusion camp say there's "no recursive self-improvement or rapid takeoff." The purest exposure — the frontier labs — is private (Anthropic, OpenAI, Cursor), so the public plays are the enablers: compute, chip-design tools, and the enterprise-agent layer.

Synthos Research · synthosresearch.com · Think Piece · grounded in ~750 tracked expert claims + our proprietary Delta signal (as of 2026-07-11) · sources below · educational only, not investment advice

What's happening at the frontier

The phrase sounds like science fiction; the mechanism is mundane. Modern AI progress runs on three inputs — writing code, running research experiments, and designing the silicon that trains the models — and AI has started to do measurable amounts of all three. No Priors flagged the first hard datapoint on June 26, 2025: DeepMind's AlphaEvolve "delivered a 23% speedup in part of Gemini's training infrastructure — early sign of AI self-improvement, though the feedback loop is months-long." Dwarkesh Patel, citing Google, noted on Feb 12, 2025 that "25% of characters checked into Google's codebase are AI-generated with human oversight." And coding has gone agentic: No Priors described Copilot's "Project Padawan" on March 13, 2025 — "assign a GitHub issue and it drafts, plans, and commits a PR like a peer programmer."

This is why our proprietary Delta signal — the rate-of-change of expert conviction — reads the way it does. One caveat up front: recursive self-improvement is not itself a standalone tracked Delta topic; the readings below are the adjacent topics that bound it. As of July 11, 2026, AI software & agents shows a net stance of +90 (on our −100…+100 scale, where +100 is unanimous bullish), "Very Bullish" — but on a thin base (n=20, now flagged low-sample), up +22 against a freshly re-distilled baseline of +68. Across those 20 recent claims: 18 bullish, 2 neutral, none bearish. The more telling shifts sit in the low-sample topics, which we flag as provisional. AI moat / disruption risk reads a net stance of +12.5 now (n=8; the model marks this "insufficient" and declines to certify a delta) against a "Very Bearish" −71 in the baseline window — an uncertified, provisional read that the panic over AI agents gutting incumbent software has faded. Frontier AI labs cooled from a +40 baseline to "Neutral / Mixed" (+19), a −21 move (small sample, n=21). And single-name NVIDIA conviction dropped a striking −77 to neutral (n=12 — small sample, provisional). Read together: conviction is migrating out of "one lab wins the intelligence explosion" and into "the loop is real, diffuse, and the enablers get paid."

All four gauges show net stance on a −100…+100 scale (bar fill = (net+100)/2). These are adjacent proxies, not a "recursive self-improvement" topic.

AI software & agents: +90 net stance
Very Bullish · +22 vs base +68 · n=20 small sample

AI moat / disruption risk: +12.5 net stance
n=8 · provisional, uncertified · was −71

Frontier AI labs: +19 net stance
Neutral · −21 vs base +40 (small sample, n=21)

AI capex / buildout: +42 net stance
Bullish · −29 vs base +71 · digesting

Delta as of 2026-07-11 · current window 2026-06-13 → 07-11 (993 claims across all tracked topics) vs. a freshly re-distilled baseline 04-18 → 06-13 (baseline total 2,870 claims). Model-extracted signal, not a Synthos price call. Recursive self-improvement is not a standalone tracked topic; the figures above are adjacent topics (AI software & agents, frontier AI labs, AI moat / disruption risk, AI capex, NVIDIA). Low-sample topics flagged provisional.

What it actually is — and why it matters

Strip the mysticism and recursive self-improvement is a set of concrete loops, each at a different stage of maturity:

1. Coding agents (live, scaling). Models write and increasingly commit code — including the code that trains the next model. No Priors on June 12, 2025: "coding excellence matters because models increasingly build the next models via systems engineering, data analysis and writing RL environments… models can now write [those RL environments], enabling further recursive self-improvement." Mark Zuckerberg, on the Dwarkesh Podcast (April 29, 2025): "within 12–18 months most code advancing Llama research will be written by goal-driven AI agents, not autocomplete."

2. AI designing chips (early, high-leverage). The tools that lay out silicon — electronic design automation, or EDA — are becoming agentic. Jensen Huang put a number on it (May 1, 2026): "the number of effective ASIC/chip designers rises ~1000x via agents, exploding Cadence's TAM for tool licenses." He's careful, though: on July 23, 2025 he said AI "will accelerate chip design and build pieces, but human creativity remains central… AI won't design [a next-gen GPU] alone."

3. Automating AI research itself (the real prize, still emerging). This is the loop that, if it closes, changes the slope. Cognitive Revolution (Aug 23, 2025): "automating AI research itself is the key accelerant — could take a lab from lead to hegemon by suddenly speeding all capability domains," scaling "from thousands to millions of automated researchers." The "AI 2027" authors Daniel Kokotajlo & Scott Alexander, on the Dwarkesh Podcast (April 3, 2025), laid out the multiplier: "a superhuman coder gives ~5x algorithmic speedup, a superhuman AI researcher ~25x, a superintelligent researcher ~hundreds-to-1000x." That is the intelligence-explosion thesis in one line — and also exactly where the disagreement lives. (Host Dwarkesh Patel argues the cautious side; see below.)

The debate: fast takeoff vs. gradual diffusion

This is a genuine, named, dated split — not a strawman. One honest wrinkle: the boldest fast-takeoff scenarios aired on the Dwarkesh Podcast came from Dwarkesh's guests — the "AI 2027" authors and Mark Zuckerberg — while host Dwarkesh Patel himself argues the cautious, gradual side. Below, each voice is labeled by who actually said it, not by whose show it aired on.

Fast takeoff — the loop compounds

"We've hit recursive self-improvement — models now build models (Codex, Opus 4.6); progress goes super-exponential, the singularity inflection point."

Jordi Visser · Feb 15, 2026 · conviction 90

Fast takeoff — coding starts it

"Coding is what starts the intelligence explosion; automating coding agents lets AI researchers speed up their own research first… AGI arrives ~2027 and superintelligence ~2028."

Daniel Kokotajlo & Scott Alexander — the "AI 2027" scenario, on the Dwarkesh Podcast · April 3, 2025 · conviction 75

Fast takeoff — the bootstrap

"Coding and AI-accelerated research are the highest-value near-term applications; models improving the next model is the bootstrap toward superintelligence."

No Priors (Sarah Guo / Elad Gil) · May 1, 2025 · conviction 85

Fast takeoff — compute demand explodes

"Agentic AI (models recursively calling themselves) raises AI compute demand by hundredfold-plus — an acceleration most underestimate because they don't code."

Forward Guidance · April 29, 2026 · conviction 85

Gradual — diffusion is a slog

"Diffusion of major technologies is always a slow slog (printing press, electricity); experts who study diffusion never see rapid takeoff, and they're right."

Dwarkesh Patel · Jan 9, 2025 · conviction 75

Gradual — the continual-learning gate

"Lack of continual learning is the core bottleneck — LLMs can't build context or improve on the job the way human employees do."

Dwarkesh Patel (his own view) · Aug 1, 2025 · conviction 80

Gradual — no takeoff, it's clustering

"No recursive self-improvement or rapid takeoff; model performance is clustering, so AI is a normal, incremental technology race, not one winner achieving AGI."

All-In · Aug 22, 2025 · conviction 80

Gradual — even Jensen hedges

"AI progress is gradual and continuous, not a single event-horizon moment where one player leaps galactically ahead."

Jensen Huang · Dec 3, 2025 · conviction 75

That continual-learning gate is Patel's own core objection — and it cuts both ways. His flip side (Aug 1, 2025): if models ever do learn on the job, "one AI learns every job… enabling a broadly deployed intelligence explosion." And No Priors, reflecting on the actual AlphaEvolve result, keeps the door honestly ajar (June 26, 2025): "whether self-improvement compounds unboundedly, plateaus, or is a one-off gain remains a genuinely open question with no answer today."

The moat problem — why "the labs win" conviction cooled

There's a second reason Frontier-labs conviction slid on our Delta: even if the loop is real, it may not be ownable. Raoul Pal (June 18, 2025): "frontier labs should be terrified of open source — two years ago comfortably behind, today only ~3 months behind the cutting edge and 90% cheaper for many use cases." Jordi Visser, the same man calling recursive self-improvement, also says (Feb 15, 2026): "Chinese open-source models (MiniMax M2.5) match Opus 4.6 at 1/20th cost… no moats." And the productivity receipts are mixed — No Priors (July 17, 2025): "enterprise coding-tool productivity impact is much lower than expected — sometimes negative or negligible — consistent with the METR report." This is the tension our founder tracks closely: if the model layer is flimsy and open-source is months behind, the durable economics sit in compute, tools and distribution, not the weights.

Where it stands, where it's going

Short (0–6 mo): coding agents keep scaling inside the labs and big-tech codebases; EDA vendors ship more agentic design flows; enterprise-agent adoption stays early — Jim Bianco (June 16, 2025) pegs it at "under 1% of enterprises using agentic AI, 80% testing… roughly 1997–98 of the dot-com analogy." Watch for a frontier-lab IPO (All-In, July 11 2026, and Anthony Pompliano, July 7 2026, both put high odds on Anthropic and/or OpenAI going public later in 2026) — the first public read on lab economics.

Medium (6–24 mo): the decisive question — does AlphaEvolve-style AI-improving-AI compound, or plateau? Dwarkesh's 12–18-month call on agent-written research code gets tested here. If the automated-research multiplier shows up in benchmark slope, fast-takeoff conviction re-rates hard; if continual learning stays unsolved, it's gradual diffusion and the enablers still win on volume.

Long (2y+): the genuine fork. Cognitive Revolution's "thousands → millions of automated researchers" world, or All-In's "clustering, normal technology race." Even the bulls disagree on timing — a guest on All-In (Sept 24, 2025) puts "superintelligent savants… in ~6–7 years via recursive self-improvement," not the 2–3 years SF insiders claim. Either way, the compute-and-power bill comes due first.

The affected map

Horizons: Short = 0–6 months · Medium = 6–24 months · Long = 2+ years. Direction chips carry words, never color alone. Tickers link to our full deep dives. The purest plays (Anthropic, OpenAI, Cursor) are private — listed for context, not as tickers.

Name	Short	Medium	Long	Why
NVDANVIDIA · our verdict: Buy — Tactical, FV ~$245	· digesting	▲ tailwind	▲▲ strong	The compute under every loop. Forward Guidance: agentic models "recursively calling themselves" raise compute demand "hundredfold-plus." Near-term our NVIDIA Delta cooled −77 to neutral (n=12 — small sample, provisional) on China-share and custom-silicon fears (OpenAI's Broadcom "jalapeno" chip, cited on the Pompliano podcast, June 27 2026); Jensen's own framing — "reinventing the computer… tens of trillions of dollars of new computers over 10 years" — is the long case.
CDNSCadence · Hold, FV ~$400	▲ tailwind	▲▲ strong	▲▲ strong	The most direct "AI-designs-chips" play. Jensen, by name: "effective ASIC/chip designers rise ~1000x via agents, exploding Cadence's TAM for tool licenses." If chip design goes agentic, EDA seat/consumption economics expand — a rare loop where the tool vendor captures the productivity gain rather than being disrupted by it.
SNPSSynopsys · Hold, FV ~$515	▲ tailwind	▲▲ strong	▲▲ strong	Same EDA duopoly logic as Cadence — agentic design flows expand the tool franchise as AI both designs chips and demands more of them. Second-order beneficiary of the 1000x-designers thesis.
GOOGLAlphabet · deep dive	▲ tailwind	▲▲ strong	· contested	Owns a live self-improvement receipt: AlphaEvolve's "23% speedup in part of Gemini's training infrastructure" (No Priors, June 26 2025) and "25% of characters checked into Google's codebase are AI-generated" (via the Dwarkesh Podcast, Feb 12 2025). Full-stack (TPUs + DeepMind + models) is a bull case; Odd Lots' "Google is gearing up to compete… technologies may end-run the current chip" cuts against NVIDIA, for Alphabet.
MSFTMicrosoft · deep dive	▲ tailwind	▲ modest	· contested	Coding-agent distribution (Copilot / "Project Padawan" commits PRs "like a peer programmer") plus OpenAI exposure. The METR caveat — enterprise coding-tool productivity "sometimes negative or negligible" — is the medium-term risk to the monetization story.
AVGOBroadcom · deep dive	▲ tailwind	▲ modest	▲ modest	The custom-silicon arm of the loop: OpenAI's Broadcom-built "jalapeno" chip "designed in 9 months" (cited on the Pompliano podcast, June 27 2026) — AI-accelerated design compressing chip cycles, and a hedge if hyperscalers route around merchant GPUs.
PLTRPalantir · Hold, FV ~$140	· neutral	▲ tailwind	▲ modest	Enterprise-agent deployment layer — where "under 1% adoption, 80% testing" (Bianco) becomes revenue if agentic workflows land in production. High expectations already in the price (our Risk 8/10); the a16z view that codified enterprise logic "isn't going anywhere" supports the moat.
NOWServiceNow · Hold	· neutral	▲ modest	· contested	Workflow incumbent monetizing agents on top of existing systems of record — Jordi Visser's "incumbent SaaS survives because users plug AI agents into existing APIs rather than rebuild from zero." Also the disruption target if agents collapse the workflow layer; a bifurcation of winners and losers.
CRMSalesforce · deep dive	· neutral	· contested	▼ risk	The clearest "will agents help or gut incumbent SaaS?" test case. a16z: "you can't replace SAP with a Postgres database plus APIs… codified business logic is the moat." The bear case: seat-based software is exactly what autonomous agents compress.
Anthropic / OpenAIfrontier labs · PRIVATE	· private	· watch	▲ the prize	The purest exposure to the loop — and you can't buy it directly. Both are rumored toward IPOs later in 2026 (All-In, July 11 2026; Pompliano, July 7 2026); OpenAI revenue rumored ~$70B this year and frontier-lab revenue "compounding well over 30%" (both All-In, July 11 2026). A listing would be the first public price on recursive-self-improvement economics. Watch, don't chase.
Cursor · open-sourcecoding tools · mostly PRIVATE	· private	· watch	· contested	Coding-agent tooling is where the loop is most visible and least investable publicly (Cursor/Anthropic/OpenAI). Raoul Pal's "open source only ~3 months behind, 90% cheaper" is the reason this layer may commoditize before anyone captures durable rent.

What we're watching (the falsifiers)

Does AlphaEvolve-style AI-improving-AI compound, or was it a one-off? No Priors called this "a genuinely open question with no answer today." A second and third self-improvement result that stacks is the single most important tell for fast takeoff.
The continual-learning gate. Dwarkesh's "core bottleneck." If models start improving on the job across deployed copies, the gradual-diffusion case weakens fast. If not, the loop stalls at "impressive autocomplete."
Enterprise coding productivity — real or AI-washing? The METR-consistent "negative or negligible" finding (No Priors, July 17 2025) vs. "Copilot already writes ~50%+ of code depending on language" (No Priors, March 13 2025). Production ROI, not demos, decides the enterprise-agent tickers.
Open-source catch-up. "~3 months behind, 90% cheaper" (Pal) and "MiniMax M2.5 at 1/20th cost" (Visser). If the gap keeps closing, moats erode and value migrates to compute/tools/distribution — the enablers, not the weights.
The first frontier-lab IPO. An Anthropic or OpenAI listing prices lab economics publicly for the first time — the market's verdict on whether recursive self-improvement is a business or a research program.

These get graded like everything else we publish, misses included — and on this topic, intellectual humility is part of the grade.

Sources

Verbatim expert claims drawn from the Synthos knowledge base (~5,000 tracked claims; ~750 on AI/agents/self-improvement): Jordi Visser (Feb 15–16, 2026), Dwarkesh Patel, his own views (Jan 9 / Aug 1, 2025); guests on the Dwarkesh Podcast — the "AI 2027" authors Daniel Kokotajlo & Scott Alexander (April 3, 2025) and Mark Zuckerberg (April 29, 2025), plus a Google datapoint cited on the show (Feb 12, 2025); No Priors — Sarah Guo & Elad Gil (March 13 / May 1 / June 12 / June 26 / July 17, 2025), Cognitive Revolution (Aug 23, 2025), Jensen Huang (July 23 / Dec 3, 2025; May 1, 2026), All-In (Aug 22 / Sept 24, 2025; July 11, 2026), Anthony Pompliano (June 27 / July 7, 2026), Forward Guidance (April 29, 2026), Raoul Pal (June 18, 2025), Jim Bianco (June 16, 2025), Odd Lots (June 18, 2025), a16z (July 7, 2026). Sentiment shifts from the Synthos Delta signal, topic-sentiment snapshot dated 2026-07-11 (adjacent topics: AI software & agents, Frontier AI labs, AI moat / disruption risk, AI capex, NVIDIA — recursive self-improvement is not itself a tracked topic). Public datapoints (AlphaEvolve 23% speedup; ~25% of characters at Google AI-generated; Copilot agent mode) as attributed by the named speakers above. Fair values and verdicts from our deep dives as of their publication dates.

Go deeper: the NVIDIA deep dive · the Cadence deep dive · the Synopsys deep dive · the Alphabet deep dive · the Palantir deep dive · the Frontier Gap Board · all our Verified Voices.

All Think Pieces Research Hub Predictions