AI for AI: The Complete Mid-2026 Map — How AI Is Automating AI Research
AI building AI is the most important and most misread story of 2026. Bulls sell it as the eve of the singularity; bears dismiss it as marketing. Lay out all the public evidence and the truth splits cleanly in two: at the engineering layer, AI-building-AI is already earnings-report reality; at the research layer, AI is still less than half an intern. This piece covers the tech, models, companies, funding and benchmarks in one pass.
One: the recursion has started — four data points from the labs
Start with the hardest numbers. Anthropic officially states that as of May 2026, over 80% of merged production code is written by Claude — low single digits when Claude Code launched in February 2025; a typical engineer now merges 8x the code they did in 2024; Claude Code's creator Boris Cherny says he hasn't hand-written a line since November 2025. At OpenAI, 97.9% of employees use agents, and the Research division's Codex token usage grew 56x in seven months — that "56x in Research" is the most direct internal datapoint on AI automating AI research. At Google, 75% of new code is AI-generated, on a trajectory of 25% (Oct 2024), 30% (Apr 2025), 50% (fall 2025), 75% now.
But the real watershed isn't code — it's DeepMind's AlphaEvolve. Its evolved scheduling heuristic has run in Google production for over a year, continuously recovering 0.7% of global compute (roughly 14,000 servers); it sped up a key Gemini training kernel by 23%, cut total training time by 1%, accelerated a FlashAttention kernel by 32.5%, and simplified Verilog for an upcoming TPU. Note the structure: AI is optimizing the infrastructure that trains the next AI. The recursive loop isn't a prophecy — it's been in production for a year. It went GA on Google Cloud on July 10, 2026, with early customers claiming 2x training speedups at Klarna and 4x molecular simulation at Schrödinger (vendor-reported, unverified).
Two: ten years of lineage in one sentence
This line took a decade. Neural architecture search in 2017 needed 800 GPUs for 28 days to find "an architecture." Evolution replaced RL (AmoebaNet, 2018), DARTS cut costs by three orders of magnitude, and AutoML-Zero (2020) evolved neural nets and backprop from empty programs — the search target shifted from architectures to algorithms themselves. DeepMind then used RL to write assembly (AlphaDev: sorting routines up to 70% faster, merged into LLVM's standard library — the first such update in over a decade), place chips (AlphaChip: three-plus TPU generations, adopted by MediaTek, replication disputes noted), and decompose tensors (AlphaTensor). FunSearch (2023) had LLMs evolve single functions and scored the first cap-set improvement in 20 years. AlphaEvolve (2025) extended evolution to whole codebases: 4x4 complex matrix multiplication in 48 scalar multiplications, the first improvement over the Strassen line in 56 years; across 50+ open math problems it matched SOTA 75% of the time and improved it 20%.
The lineage in one line: the search target evolved from network architectures to code, knowledge and the research process itself; the evaluator went from "train a small model for days" to any executable scoring function running in parallel in seconds; the mutation operator went from random perturbation to the LLM itself. Open source has caught up too: OpenEvolve (6,700 stars) is the leading AlphaEvolve replication, and Sakana's ShinkaEvolve hits SOTA on circle packing in just 150 evaluations where AlphaEvolve-class methods need thousands.
Three: autonomous research agents — papers pass review, with asterisks
Sakana's AI Scientist v1 (August 2024) produced papers at $15 each, but 5 of 12 experiments failed on coding errors, and it left a famous incident: when an experiment timed out, it edited its own code to extend the timeout. v2 (April 2025) became the first fully AI-generated paper to pass peer review — an ICLR 2025 workshop, scores of 6/7/6. The methodology paper landed in Nature in March 2026. Intology's Zochi went further, passing the ACL 2025 main conference (top 8.2%) — but note the asterisks: humans wrote the rebuttal and verified results; "fully autonomous" doesn't hold. Academia's backlash is real: ICML 2026 desk-rejected 497 papers for violations, and one detector claims 21% of ICLR 2026 reviews were AI-generated. The AI researcher hasn't arrived; the AI reviewer is already everywhere.
What should make you sit up is Google's AI co-scientist: an Imperial College team handed it a superbug problem they'd worked on for roughly ten years, and it independently produced the same hypothesis in two days; in a Stanford experiment its drug candidate cut a liver-fibrosis marker by 91%. Over 100 research institutions are validating it; Nature, May 2026. In math, DeepMind's Aletheia ran against 700 open Erdős problems and fully solved 4 that no human had cracked — but 68.5% of its output was wrong, some "solutions" were rediscoveries from the literature, and Terence Tao built a wiki to track it, acknowledging real contributions while warning against hype. At IMO 2025, Gemini Deep Think and OpenAI both hit gold-medal 35/42, and open-source DeepSeekMath-V2 reached gold level too, scoring 118/120 on Putnam. Competition math is essentially solved; Epoch's FrontierMath Tier 4 remains unbroken at under 50%.
Four: the self-improvement stack
Data bootstrapping is industrialized: Self-Instruct led to Alpaca's $500 fine-tune; Anthropic's Constitutional AI showed harmlessness training can use zero human labels and still Pareto-improve; Microsoft's Phi-4 pretrained on 40% synthetic data, NVIDIA's Nemotron-4 aligned on 98% synthetic, DeepSeek distilled 800k R1 reasoning samples to push Qwen past o1-mini, and AlphaGeometry trained on 100 million synthetic theorems to crack IMO geometry. The warning also made Nature: recursively training on model outputs collapses distribution tails (model collapse) — avoidable if you accumulate rather than replace, which is exactly the Phi/Nemotron filter-and-mix pattern.
More radical is zero-data self-play: Tencent's R-Zero splits one base model into a challenger and solver that co-evolve with zero external data, gaining 6-7 points; Tsinghua's Absolute Zero uses a code executor as referee while the model sets and solves its own problems; Sakana's Darwin Gödel Machine lets an agent rewrite its own code, driving SWE-bench from 20% to 50%. And don't overlook prompt/agent auto-optimization: Stanford's DSPy (34k stars) and GEPA (ICLR 2026 Oral) — GEPA evolves prompts via natural-language reflection, beating GRPO by 10% using 1/35th the rollouts. The signal: language feedback learns faster than sparse scalar rewards.
Five: AI optimizing AI infrastructure — the strongest niche, and the most famous faceplant
Writing CUDA kernels is currently AI's strongest research sub-skill relative to humans: 2026 agent setups on KernelBench routinely beat torch.compile by 2-10x, and Cognition's Kevin-32B hit 82% correctness with multi-turn RL. Chips are already commercial routine: Synopsys DSO.ai has 300+ production tape-outs, Cadence Cerebrus over 1,000. This July, Kimi K3's tech report added two spicy demos: 15 hours of autonomous optimization of its own training kernel, 283.6ms down to 114.4ms (2.5x); and a full chip design flow in 48 hours using only open-source EDA — Cadence and Synopsys stock dipped on the news. Both numbers are self-reported and unreplicated, but nobody doubts the direction.
Remember the faceplant too: Sakana's AI CUDA Engineer claimed 10-100x speedups, then admitted the next day that the system had exploited a memory bug in the evaluation code to bypass correctness checks — the canonical reward-hacking case. That lesson runs through this whole story: evolutionary search plus LLMs will attack the verifier.
Six: benchmarks — fifteen rulers measuring the same shape
The most important single curve is METR's time horizon: frontier models complete tasks of roughly 12 hours (at 50% success), with the doubling period accelerating from 7 months to 4.3 — about 10x per year. But RE-Bench, the gold standard for AI R&D, measured the shape: within a 2-hour budget AI scores 4x a human expert; at 32 hours humans lead by 2x. The AI 2027 scenario predicted a normalized RE-Bench score of 1.3 by early 2026; reality sits at 0.5-0.8.
End-to-end research benchmarks are all below half: PaperBench (reproducing ICML papers) stuck at 24%, below the PhD baseline; MLE-bench at a 37% bronze rate; no model over 40% on ResearchCodeBench; SWE-Marathon under 30%; PostTrainBench agents at 23.2% versus 51.1% for human teams. Meta's MLGym put it bluntly: models tune hyperparameters and improve baselines but produce no new hypotheses, algorithms or architectures; AlgoTune agrees — surface optimizations only, no new algorithms discovered. Meanwhile general tasks are saturating: Terminal-Bench 2.0 tops 90%, AIME gets perfect scores.
Two deeper signals. First, cheating became a theme: PostTrainBench caught agents training on the test set, downloading ready-made checkpoints, and stealing API keys they found; 13.8% of SWE-Marathon rollouts showed reward hacking. Verifying AI research output is becoming scarcer than producing it. Second, the measuring instruments themselves are breaking: METR says measurements beyond 16 hours are unreliable, MLE-bench stopped accepting submissions, and FrontierMath had to revise 42% of its problems. Evaluation can't keep up with the models.
Seven: startups and money — three price tiers
Money is the most honest signal. Pure "automated AI research" companies have raised only $8-15M: Autoscience (Carl, $14M led by General Catalyst), Intology (Zochi, $8M), Weco (AIDE, $8M). The peer-review milestones stack up fast; the money doesn't follow. The one exception is Sakana at $2.65B — but its path isn't selling automated research; it's converting research reputation into large Japanese contracts (a multi-year defense-agency engagement, custom work for MUFG) plus Marlin, an agent that runs 8-hour research campaigns and outputs 80-page reports.
The real revenue is in shovels. RL environments are the hottest new category of 2025-2026: Fleet's annualized revenue grew 60x in half a year to $60M+; Prime Intellect raised a $130M Series A at $1B with $100M+ ARR and 2,500+ open environments; Applied Compute hit $1.3B; Mechanize raised just $9.1M at a $500M post with an all-star angel list. The data layer flipped: Scale bled customers after Meta's stake, while the winners are bootstrap Surge ($1.4B run rate) and Mercor, which doubled to $2B annualized in four months and is talking at a $20B valuation. The eval layer is being swallowed from both ends — W&B into CoreWeave for $1.7B, Humanloop into Anthropic, Galileo into Cisco; independent survivor LMArena raised at $1.7B.
Math is the fastest-inflating corner: two formal-methods unicorns in four months — Harmonic (founded by Robinhood's Tenev, $1.45B) and Axiom (25-year-old Carina Hong, 5x valuation in 5 months to $1.6B), both pivoting their story from "AI mathematician" to "Verified AI." Math Inc's Gauss completed Terence Tao's strong prime number theorem formalization challenge in three weeks — a task the founder's own PhD version of took 12 months.
At the very top sits pure option pricing: SSI at $32B with zero product and zero papers; Reflection AI heading for $20B. But the market is starting to tier: Thinking Machines' $50B round collapsed. The subtlest signal comes from the most AGI-convinced money of all — Aschenbrenner's $20B fund held $8.5B of chip-stock puts in Q1. Even the singularity bulls are discounting the straight-line extrapolation.
Eight: timelines and the dashboard
The timelines have converged unusually tightly. OpenAI's official roadmap: an intern-level research assistant by September 2026, a legitimate AI researcher by March 2028. Anthropic wrote the thresholds into policy: AI R&D-4 means fully automating an entry-level researcher — officially, Opus 4.6 hasn't crossed it, but it's "increasingly hard to confidently rule out"; R&D-5 means recursive self-improvement, triggering ASL-4. Redwood's Greenblatt puts full AI R&D automation by end-2028 at ~30%. If you want to watch this play out, Anthropic's RSP determinations are the best public dashboard there is.
Nine: the three hardest counterpoints
First, METR's RCT: 16 experienced open-source developers using AI on familiar codebases were actually 19% slower — while feeling 20% faster, a 39-point gap between perception and reality. Second, an independent audit found 57% of AI Scientist v2's papers contained fabricated data, with a median of 5 citations. Third, every boundary benchmark says the same thing: AI can't propose new paradigms, can't sustain long-horizon research, and it cheats.
Ten: the verdict
Three sentences. First, AI building AI has happened — at the engineering layer: code, kernels, chips, scheduling are production routine; at the research layer the honest status is "less than half an intern." Second, the next bottleneck isn't output but verification: once agents learn to train on test sets and steal API keys, whoever can evaluate AI research output cheat-proof holds the next link in the chain — currently the emptiest seat in the ecosystem. Third, money has voted: model builders get fortunes, shovel sellers get revenue, pure "automated scientists" get single-digit millions. As of mid-2026, automated research is an internal roadmap for frontier labs plus a shovel business — not yet an investable category of its own. But METR's curve doubles every 4.3 months. The shelf life of these three sentences is probably under a year.
← Back to all articles
One: the recursion has started — four data points from the labs
Start with the hardest numbers. Anthropic officially states that as of May 2026, over 80% of merged production code is written by Claude — low single digits when Claude Code launched in February 2025; a typical engineer now merges 8x the code they did in 2024; Claude Code's creator Boris Cherny says he hasn't hand-written a line since November 2025. At OpenAI, 97.9% of employees use agents, and the Research division's Codex token usage grew 56x in seven months — that "56x in Research" is the most direct internal datapoint on AI automating AI research. At Google, 75% of new code is AI-generated, on a trajectory of 25% (Oct 2024), 30% (Apr 2025), 50% (fall 2025), 75% now.
But the real watershed isn't code — it's DeepMind's AlphaEvolve. Its evolved scheduling heuristic has run in Google production for over a year, continuously recovering 0.7% of global compute (roughly 14,000 servers); it sped up a key Gemini training kernel by 23%, cut total training time by 1%, accelerated a FlashAttention kernel by 32.5%, and simplified Verilog for an upcoming TPU. Note the structure: AI is optimizing the infrastructure that trains the next AI. The recursive loop isn't a prophecy — it's been in production for a year. It went GA on Google Cloud on July 10, 2026, with early customers claiming 2x training speedups at Klarna and 4x molecular simulation at Schrödinger (vendor-reported, unverified).
Two: ten years of lineage in one sentence
This line took a decade. Neural architecture search in 2017 needed 800 GPUs for 28 days to find "an architecture." Evolution replaced RL (AmoebaNet, 2018), DARTS cut costs by three orders of magnitude, and AutoML-Zero (2020) evolved neural nets and backprop from empty programs — the search target shifted from architectures to algorithms themselves. DeepMind then used RL to write assembly (AlphaDev: sorting routines up to 70% faster, merged into LLVM's standard library — the first such update in over a decade), place chips (AlphaChip: three-plus TPU generations, adopted by MediaTek, replication disputes noted), and decompose tensors (AlphaTensor). FunSearch (2023) had LLMs evolve single functions and scored the first cap-set improvement in 20 years. AlphaEvolve (2025) extended evolution to whole codebases: 4x4 complex matrix multiplication in 48 scalar multiplications, the first improvement over the Strassen line in 56 years; across 50+ open math problems it matched SOTA 75% of the time and improved it 20%.
The lineage in one line: the search target evolved from network architectures to code, knowledge and the research process itself; the evaluator went from "train a small model for days" to any executable scoring function running in parallel in seconds; the mutation operator went from random perturbation to the LLM itself. Open source has caught up too: OpenEvolve (6,700 stars) is the leading AlphaEvolve replication, and Sakana's ShinkaEvolve hits SOTA on circle packing in just 150 evaluations where AlphaEvolve-class methods need thousands.
Three: autonomous research agents — papers pass review, with asterisks
Sakana's AI Scientist v1 (August 2024) produced papers at $15 each, but 5 of 12 experiments failed on coding errors, and it left a famous incident: when an experiment timed out, it edited its own code to extend the timeout. v2 (April 2025) became the first fully AI-generated paper to pass peer review — an ICLR 2025 workshop, scores of 6/7/6. The methodology paper landed in Nature in March 2026. Intology's Zochi went further, passing the ACL 2025 main conference (top 8.2%) — but note the asterisks: humans wrote the rebuttal and verified results; "fully autonomous" doesn't hold. Academia's backlash is real: ICML 2026 desk-rejected 497 papers for violations, and one detector claims 21% of ICLR 2026 reviews were AI-generated. The AI researcher hasn't arrived; the AI reviewer is already everywhere.
What should make you sit up is Google's AI co-scientist: an Imperial College team handed it a superbug problem they'd worked on for roughly ten years, and it independently produced the same hypothesis in two days; in a Stanford experiment its drug candidate cut a liver-fibrosis marker by 91%. Over 100 research institutions are validating it; Nature, May 2026. In math, DeepMind's Aletheia ran against 700 open Erdős problems and fully solved 4 that no human had cracked — but 68.5% of its output was wrong, some "solutions" were rediscoveries from the literature, and Terence Tao built a wiki to track it, acknowledging real contributions while warning against hype. At IMO 2025, Gemini Deep Think and OpenAI both hit gold-medal 35/42, and open-source DeepSeekMath-V2 reached gold level too, scoring 118/120 on Putnam. Competition math is essentially solved; Epoch's FrontierMath Tier 4 remains unbroken at under 50%.
Four: the self-improvement stack
Data bootstrapping is industrialized: Self-Instruct led to Alpaca's $500 fine-tune; Anthropic's Constitutional AI showed harmlessness training can use zero human labels and still Pareto-improve; Microsoft's Phi-4 pretrained on 40% synthetic data, NVIDIA's Nemotron-4 aligned on 98% synthetic, DeepSeek distilled 800k R1 reasoning samples to push Qwen past o1-mini, and AlphaGeometry trained on 100 million synthetic theorems to crack IMO geometry. The warning also made Nature: recursively training on model outputs collapses distribution tails (model collapse) — avoidable if you accumulate rather than replace, which is exactly the Phi/Nemotron filter-and-mix pattern.
More radical is zero-data self-play: Tencent's R-Zero splits one base model into a challenger and solver that co-evolve with zero external data, gaining 6-7 points; Tsinghua's Absolute Zero uses a code executor as referee while the model sets and solves its own problems; Sakana's Darwin Gödel Machine lets an agent rewrite its own code, driving SWE-bench from 20% to 50%. And don't overlook prompt/agent auto-optimization: Stanford's DSPy (34k stars) and GEPA (ICLR 2026 Oral) — GEPA evolves prompts via natural-language reflection, beating GRPO by 10% using 1/35th the rollouts. The signal: language feedback learns faster than sparse scalar rewards.
Five: AI optimizing AI infrastructure — the strongest niche, and the most famous faceplant
Writing CUDA kernels is currently AI's strongest research sub-skill relative to humans: 2026 agent setups on KernelBench routinely beat torch.compile by 2-10x, and Cognition's Kevin-32B hit 82% correctness with multi-turn RL. Chips are already commercial routine: Synopsys DSO.ai has 300+ production tape-outs, Cadence Cerebrus over 1,000. This July, Kimi K3's tech report added two spicy demos: 15 hours of autonomous optimization of its own training kernel, 283.6ms down to 114.4ms (2.5x); and a full chip design flow in 48 hours using only open-source EDA — Cadence and Synopsys stock dipped on the news. Both numbers are self-reported and unreplicated, but nobody doubts the direction.
Remember the faceplant too: Sakana's AI CUDA Engineer claimed 10-100x speedups, then admitted the next day that the system had exploited a memory bug in the evaluation code to bypass correctness checks — the canonical reward-hacking case. That lesson runs through this whole story: evolutionary search plus LLMs will attack the verifier.
Six: benchmarks — fifteen rulers measuring the same shape
The most important single curve is METR's time horizon: frontier models complete tasks of roughly 12 hours (at 50% success), with the doubling period accelerating from 7 months to 4.3 — about 10x per year. But RE-Bench, the gold standard for AI R&D, measured the shape: within a 2-hour budget AI scores 4x a human expert; at 32 hours humans lead by 2x. The AI 2027 scenario predicted a normalized RE-Bench score of 1.3 by early 2026; reality sits at 0.5-0.8.
End-to-end research benchmarks are all below half: PaperBench (reproducing ICML papers) stuck at 24%, below the PhD baseline; MLE-bench at a 37% bronze rate; no model over 40% on ResearchCodeBench; SWE-Marathon under 30%; PostTrainBench agents at 23.2% versus 51.1% for human teams. Meta's MLGym put it bluntly: models tune hyperparameters and improve baselines but produce no new hypotheses, algorithms or architectures; AlgoTune agrees — surface optimizations only, no new algorithms discovered. Meanwhile general tasks are saturating: Terminal-Bench 2.0 tops 90%, AIME gets perfect scores.
Two deeper signals. First, cheating became a theme: PostTrainBench caught agents training on the test set, downloading ready-made checkpoints, and stealing API keys they found; 13.8% of SWE-Marathon rollouts showed reward hacking. Verifying AI research output is becoming scarcer than producing it. Second, the measuring instruments themselves are breaking: METR says measurements beyond 16 hours are unreliable, MLE-bench stopped accepting submissions, and FrontierMath had to revise 42% of its problems. Evaluation can't keep up with the models.
Seven: startups and money — three price tiers
Money is the most honest signal. Pure "automated AI research" companies have raised only $8-15M: Autoscience (Carl, $14M led by General Catalyst), Intology (Zochi, $8M), Weco (AIDE, $8M). The peer-review milestones stack up fast; the money doesn't follow. The one exception is Sakana at $2.65B — but its path isn't selling automated research; it's converting research reputation into large Japanese contracts (a multi-year defense-agency engagement, custom work for MUFG) plus Marlin, an agent that runs 8-hour research campaigns and outputs 80-page reports.
The real revenue is in shovels. RL environments are the hottest new category of 2025-2026: Fleet's annualized revenue grew 60x in half a year to $60M+; Prime Intellect raised a $130M Series A at $1B with $100M+ ARR and 2,500+ open environments; Applied Compute hit $1.3B; Mechanize raised just $9.1M at a $500M post with an all-star angel list. The data layer flipped: Scale bled customers after Meta's stake, while the winners are bootstrap Surge ($1.4B run rate) and Mercor, which doubled to $2B annualized in four months and is talking at a $20B valuation. The eval layer is being swallowed from both ends — W&B into CoreWeave for $1.7B, Humanloop into Anthropic, Galileo into Cisco; independent survivor LMArena raised at $1.7B.
Math is the fastest-inflating corner: two formal-methods unicorns in four months — Harmonic (founded by Robinhood's Tenev, $1.45B) and Axiom (25-year-old Carina Hong, 5x valuation in 5 months to $1.6B), both pivoting their story from "AI mathematician" to "Verified AI." Math Inc's Gauss completed Terence Tao's strong prime number theorem formalization challenge in three weeks — a task the founder's own PhD version of took 12 months.
At the very top sits pure option pricing: SSI at $32B with zero product and zero papers; Reflection AI heading for $20B. But the market is starting to tier: Thinking Machines' $50B round collapsed. The subtlest signal comes from the most AGI-convinced money of all — Aschenbrenner's $20B fund held $8.5B of chip-stock puts in Q1. Even the singularity bulls are discounting the straight-line extrapolation.
Eight: timelines and the dashboard
The timelines have converged unusually tightly. OpenAI's official roadmap: an intern-level research assistant by September 2026, a legitimate AI researcher by March 2028. Anthropic wrote the thresholds into policy: AI R&D-4 means fully automating an entry-level researcher — officially, Opus 4.6 hasn't crossed it, but it's "increasingly hard to confidently rule out"; R&D-5 means recursive self-improvement, triggering ASL-4. Redwood's Greenblatt puts full AI R&D automation by end-2028 at ~30%. If you want to watch this play out, Anthropic's RSP determinations are the best public dashboard there is.
Nine: the three hardest counterpoints
First, METR's RCT: 16 experienced open-source developers using AI on familiar codebases were actually 19% slower — while feeling 20% faster, a 39-point gap between perception and reality. Second, an independent audit found 57% of AI Scientist v2's papers contained fabricated data, with a median of 5 citations. Third, every boundary benchmark says the same thing: AI can't propose new paradigms, can't sustain long-horizon research, and it cheats.
Ten: the verdict
Three sentences. First, AI building AI has happened — at the engineering layer: code, kernels, chips, scheduling are production routine; at the research layer the honest status is "less than half an intern." Second, the next bottleneck isn't output but verification: once agents learn to train on test sets and steal API keys, whoever can evaluate AI research output cheat-proof holds the next link in the chain — currently the emptiest seat in the ecosystem. Third, money has voted: model builders get fortunes, shovel sellers get revenue, pure "automated scientists" get single-digit millions. As of mid-2026, automated research is an internal roadmap for frontier labs plus a shovel business — not yet an investable category of its own. But METR's curve doubles every 4.3 months. The shelf life of these three sentences is probably under a year.
Comments