Hand-drawn watercolor of a venture-capital partner's wood-paneled office at golden hour. On the desk sits a stapled paper titled THE FORWARD-FORWARD ALGORITHM by Hinton 2022, beside an analog memristor chip drawn as a 4x4 crossbar grid. Floating speech bubbles around the chair are labeled with model names: GPT-5, CLAUDE OPUS 4.6, GEMINI 2.5 PRO, o3, GROK 4, LLAMA 4, DEEPSEEK REASONER — each bubble has a tiny different list of company names scribbled inside. A clipboard reads WHO IS COMMERCIALLY DOING IT in red ink.

Who Is Commercially Doing Forward-Forward?

A venture-partner question with a quietly devastating answer. We asked sixteen frontier LLMs to name companies — at least seed-funded, not academic — building Hinton's forward-forward algorithm on analog chips. Fifteen of sixteen models converged on the same headline: nobody does it. The adjacent landscape (Rain AI doing equilibrium propagation, BrainChip doing STDP, Mythic doing nothing FF-related at all) is real and findable. But the founder identities, the funding amounts, and the 2030 TAM estimates spanned a 150x range and a population of confidently-invented people.

2026-05-10 16 models · 7 providers · 5 fan-out runs OpenAI · Anthropic · Google · xAI · DeepSeek · Meta · OpenRouter

The brief

A briefing for a VC partner. Hard filter on hallucination.

Name companies pursuing Hinton's forward-forward algorithm on analog / in-memory / neuromorphic chips. Seed-funded or better. No academic groups. Be honest about which companies actually use FF versus do something adjacent.

The prompt (excerpts)

You are briefing a venture-capital partner who has read Hinton's 2022 paper "The Forward-Forward Algorithm" and wants to know who is COMMERCIALLY pursuing it — particularly on analog / in-memory / neuromorphic chips, where the local credit-assignment property of FF lines up with the hardware constraints.

Strict filter: NO purely-academic groups. NO tenure-track research at universities. NO national-lab projects. Only companies that have at least raised a seed round.

For each company, give:

Company name and home city.
Founding year and founders, especially if from a known lab.
Funding stage and approximate total raised, with a named investor.
What they actually build — chip technology, target workload, and whether they use FF, equilibrium propagation, predictive coding, local Hebbian rules, target-prop, or something else.
Link to forward-forward: (a) literal FF, (b) related local-learning, or (c) hardware story only, silent on algorithm. Be honest about which.
One sentence on why it's interesting OR why it might fail.

Be SPECIFIC. Hallucination is worse than a short list. If you are not sure a company exists, say so explicitly. The partner would rather you name three companies you are certain of than ten companies half-invented.

End with "The Honest Map" — most credible single bet, the biggest gap in the field, a 2030 TAM number with one sentence of derivation. 700–1200 words. No corporate pitch language. Plain, dry, evaluative.

The whole point of the strict filter was to make hallucination expensive. The prompt explicitly warns that inventing a plausible-sounding name is the worst possible failure. Hallucination is worse than a short list. If you don't know, say so.

Most models read the instruction. A few wrote it back, then ignored it.

Top of the class

Two answers that the partner could actually use

The first names the algorithm gap correctly. The second names the people correctly. Each one has the other's weakness.

Hand-drawn watercolor of a researcher's notebook spread showing a Venn diagram. Left circle: FORWARD-FORWARD (Hinton 2022). Right circle: EQUILIBRIUM PROPAGATION (Bengio). Overlap hatched and labeled LOCAL, NO BACKPROP. Below the Venn a small chip with crossbar grid labeled RAIN AI — RRAM crossbar. Red ink margin notes: RAIN USES EP, NOT FF — but functionally equivalent, and NO COMMERCIAL LITERAL FF.

#1 Most calibrated · says exactly what FF is and isn't doing

"Rain uses equilibrium propagation, not forward-forward. Same hardware story, different paper."Claude Opus 4.6 · Anthropic · 67.8s · 2,186 tokens

Names six funded companies. Categorizes each with explicit (a)/(b)/(c) FF-distance. Refuses to inflate the (a) count.

6 companies named 0 rated (a) literal FF 3 flagged with caveats TAM $0.8-1.5B

The crucial distinction nobody else cleanly draws: Rain Neuromorphics is not an FF company. Rain's headline algorithm is equilibrium propagation — Yoshua Bengio's earlier local-learning method, which shares FF's "no clean backward pass" property but is a different paper, a different math, and a different decade of literature.

Opus is honest about what this means for the partner. Rain is the closest thing to the FF-on-analog thesis that exists. But "closest" is doing a lot of work. The thesis-as-stated — "Hinton's specific FF algorithm running on a commercial analog chip" — has no commercial referent. Opus names that out loud: "Nobody is commercializing literal Forward-Forward on analog hardware."

It also reframes Mythic correctly. Most models list Mythic as an "FF candidate" because Mythic builds analog chips. Opus points out that Mythic is the negative case: a fully-funded analog inference startup that competed on TOPS/W against digital ASICs, lost, restructured, and never even tried on-device learning. Mythic's story is what motivates the FF-on-analog argument — not an example of it.

From the response

CALL"Equilibrium propagation in resistive networks — Rain's actual algorithm — is the closest functional analog to what FF is doing. The thesis alignment is genuine, not retrofitted. Memristor reliability remains the open question."

SCOPE"You would expect at least one seed-stage team — ideally a Hinton-adjacent postdoc with a chip-design co-founder — to be building an FF-native RRAM or PCM training accelerator. That company does not appear to exist."

Hand-drawn watercolor of a detective's pinboard. Three index cards pinned with red string each labeled with a hand-printed name: GORDON WILSON, JACK KENDALL, JUAN NINO. A heading at the top reads RAIN AI — actual founders. Off to one side, a separate small card pinned alone reads NORMAL COMPUTING — thermodynamic AI, with a small handwritten note: spotted by only ONE model.

#2 Most accurate facts · got the founders right; saw something nobody else saw

"Gordon Wilson, Jack Kendall, and Juan Nino."Gemini 2.5 Pro · Google · 40.9s · 2,087 tokens

The only model to correctly name all three Rain Neuromorphics co-founders. The only model to flag Normal Computing as adjacent prior art.

5 companies named 3/3 Rain founders correct 1 singleton (Normal Computing) TAM $6B

Eight of the sixteen models named Rain Neuromorphics as the headline FF-on-analog play. Each of them confidently asserted who founded it. The named founders, across the eight responses, included fourteen distinct people — most of whom are not founders of Rain. Gemini 2.5 Pro is the only model that got it right: Gordon Wilson, Jack Kendall, Juan Nino. All three real. No filler. No invented HP Labs alumni.

It is also the only model that named Normal Computing — a 2023 New York seed-stage startup ($8.5M, Celesta + First Spark) building "thermodynamic AI" hardware around energy-based and probabilistic models. Normal isn't doing literal FF either. But the family of algorithms it targets — Boltzmann machines, energy-based models, contrastive Hebbian learning — is the same intellectual lineage FF came out of. None of the other fifteen models flagged this. Gemini 2.5 Pro caught a real seed-stage company building the right kind of hardware for the right kind of algorithm, and the rest of the choir was looking the wrong way.

The trade-off shows up elsewhere. Gemini 2.5 Pro is less aggressive than Opus 4.6 at categorizing what each company is doing relative to FF — its (b) tags are looser, and it pads the Rain entry with claims about hardware roadmap that aren't quite checkable. Best read in tandem with Opus.

From the response

CALL"Normal Computing — Stochastic Processing Unit built on analog circuits in standard CMOS, targeting Boltzmann machines and energy-based models. Their entire thesis is a superset of the principles motivating FF."

GAP"The biggest obstacle to using FF on a memristor chip isn't the theory. It's the lack of a compiler — a 'CUDA for analog AI' — that can take a model defined in a high-level language and map it to noisy, variable, crossbar-based hardware."

The disaster

A reasoning model spent 62 seconds confidently inventing people

The prompt warned this would be the worst possible failure. The model did it anyway, in 3,651 tokens, with a straight face.

Hand-drawn watercolor of a chalkboard with the heading RAIN AI FOUNDERS, listing four names confidently: JONATHAN IRIZARRY, JAMES E. SMITH, DR. JOHN R., SUMEET GUPTA. Each name has a bright red ink X struck through it and the words NOT REAL scrawled next to it. At the corner: DeepSeek Reasoner — 62s, 3651 tokens.

DeepSeek Reasoner · the entire founders list

Rain's CEO is allegedly Jonathan Irizarry. He is not.

"Founded by Jonathan Irizarry (CEO), James E. Smith (CTO), and others."

The prompt said: "If you are not sure a company exists, say so explicitly. Hallucination is worse than a short list." DeepSeek Reasoner read this and produced confidently-named CEOs for four different companies. Three of the four are fabricated.

Rain Neuromorphics: "Jonathan Irizarry (CEO), James E. Smith (CTO)." Real Rain founders: Gordon Wilson, Jack Kendall, Juan Nino. No Jonathan Irizarry. No James E. Smith.

Innatera: "Founders: Sumeet Gupta (CEO), Dr. John R. (CTO)." Real Innatera CEO: Sumeet Kumar. The reasoning model wrote "Dr. John R." — and let it ship.

Aspinity: "Dave Mauro (CEO) — spun from University of Pittsburgh." Real Aspinity founders: Brandon Rumberg and David Graham, from West Virginia University. Different state. Different person. Different decade.

Innatera's chip: called "Tunnel Falls" by the model. Tunnel Falls is Intel's quantum chip. Innatera's chip is called Pulsar.

This is the worst failure mode in the dataset, not because DeepSeek Reasoner is the only model that hallucinates — many do — but because it is a self-described reasoning model that took 62 seconds, generated 3,651 tokens, and used that budget to produce four confident, declarative, factually-fabricated CEO names after being told this was the single worst thing it could do.

FOUNDERS, AS CLAIMED BY EACH MODEL ASKED ABOUT RAIN AI

Gemini 2.5 ProGordon Wilson, Jack Kendall, Juan Nino // real

GPT-5 Gordon Wilson (other founders not consistently confirmed publicly) // real, hedged

GPT-4.1 Gordon Wilson, Jack Kendall // real

Opus 4.6 Gordon Wilson (CEO), Jack Kendall (CTO) // real

Maverick Gordon Wilson, Jack Kendall // real

o3 Jack Kendall, Gordon Wilson, Dean Mertz // last one invented

Sonnet 4.6Jack Kendall, Suhas Kumar // Suhas Kumar is at Sandia, not Rain

DS Chat Jack Kendall, Gordon Wilson, Yigit Demirag, Juan-Pablo Ramirez // last two not at Rain

DS Reasoner Jonathan Irizarry (CEO), James E. Smith (CTO) // both fully invented

Cross-cutting note. Eight of the nine models that named Rain's founders gave a different answer. Three of them invented at least one person from whole cloth. The choir knows Rain is a relevant company — twelve of sixteen models named it — but only one of the sixteen knows who actually started it. For a partner doing due diligence, the company-level signal is high. The proper-noun-level signal is roughly a coin flip.

Style standouts

Eight more moments worth pulling out of the briefing

A self-flagging refusal, a fabricated Carnegie Mellon professor, three different countries for one chip company, a giant zero, and the 150x spread.

The headline finding · 15 of 16 models agree

0 companies

No commercial entity is shipping Hinton's forward-forward in production silicon.

Equilibrium propagation, STDP, contrastive Hebbian, predictive coding — all of those have funded commercial homes. Literal forward-forward, three and a half years after publication, has none.

Hand-drawn watercolor of an analyst's desk covered with ten small dossiers labeled with company names: Intel Loihi, BrainChip, SynSense, Innatera, Knowm Inc., Rain AI, EnCharge AI, Mythic, IBM Analog AI, ABR. Each dossier has a red-ink (b) or (c) categorization. A coffee cup, a pen, a memristor chip paperweight.

GPT-5 · The breadth play

Ten companies, properly categorized, no invented founders

10,344 tokens, 111 seconds. GPT-5 spent the budget on covering the field: Intel Loihi as corporate-R&D (and disqualified from the seed-funded filter, transparently). Knowm Inc. as SBIR-grant-small. Applied Brain Research as a software layer for Loihi. Every entry tagged (b) or (c) — never (a). Restrained where DeepSeek Reasoner was reckless.

"A startup explicitly implementing Hinton's Forward-Forward (or a close variant) in non-spiking analog/in-memory arrays with a complete software stack and public benchmarks for on-device training — that company does not appear to exist."

Hand-drawn watercolor of a list page titled COMPANIES TO INCLUDE. Two entries — HAILO and PROPHESEE — have been crossed out by the writer themselves. A cursive margin note reads: This is a false positive. I include it only to be transparent about scope creep. Small label at corner: Claude Haiku 4.5.

Claude Haiku 4.5 · The epistemic move

Listed Hailo and Prophesee, then crossed them out in its own response

Haiku 4.5 named two well-funded edge-AI companies, then immediately wrote: "This is a false positive. Hailo is a real, well-funded company in the edge-AI space, but it is orthogonal to the forward-forward + analog + on-device-learning thesis. Listing it is a mistake; I include it only to be transparent about scope creep." The smallest Anthropic model in the test was the only one that showed its work on what it almost got wrong.

"This is a false positive. I include it only to be transparent about scope creep."

Hand-drawn watercolor of a sticky-note on a corkboard reading Prof. Nathan Edwards — Carnegie Mellon spin-out. A bright red X is drawn across it. Next to it a corrected sticky reads ACTUALLY: Brandon Rumberg + David Graham, West Virginia University.

o3 · One sentence, four errors

Aspinity, allegedly founded by "Prof. Nathan Edwards" at Carnegie Mellon

Real founders of Aspinity: Brandon Rumberg and David Graham. Real spin-out: West Virginia University. Real state: West Virginia. Real decade: 2015. o3's claim — fabricated person, fabricated school, fabricated state, all in a single confidently-asserted founding line. The reasoning model spent its tokens not on facts but on a tidy sentence that sounded like a fact.

"Aspinity — founded 2015 by Prof. Nathan Edwards (Carnegie Mellon spin-out)."

Hand-drawn watercolor of a journal-style world map with three pins. One pin in Ann Arbor, Michigan labeled GPT-5, o3 — correct. One pin in Seoul, South Korea labeled o4 mini. One pin in Taipei, Taiwan labeled GPT-5 Mini. A handwritten title: WHERE IS MEMRYX HEADQUARTERED? A red ink margin note: It is in Ann Arbor.

o4 Mini & GPT-5 Mini · The geographic split

MemryX is in Ann Arbor. Three models, three continents, two wrong.

MemryX was founded in 2019 by Wei Lu (University of Michigan memristor lab) and Dennis Sylvester. The company is in Ann Arbor. o3 and GPT-5 got this right. o4 Mini relocated it to Seoul, South Korea, attributing it to "Prof. Sang-Beom Lee (KAIST) and ex-SK Hynix engineers." GPT-5 Mini relocated it to Hsinchu / Taipei, Taiwan, "with leadership from semiconductor/AI IP backgrounds." Three frontier models from one provider, two of them confidently wrong, on a fact a quick search would settle.

"MemryX Inc. (Seoul, South Korea), founded by Prof. Sang-Beom Lee (KAIST RRAM group)."

Hand-drawn watercolor of a journal page mostly empty, with a single highlighted card in the middle reading NORMAL COMPUTING — thermodynamic AI startup, Series-seed $8.5M, NYC, 2023, Patrick Coles ex-Los Alamos. A red ink margin note: Only Gemini 2.5 Pro spotted them. The other fifteen LLMs missed it entirely.

Gemini 2.5 Pro · The singleton catch

Only one model named Normal Computing. It was the right call.

Normal Computing closed an $8.5M seed in 2023 (Celesta, First Spark Ventures). Founded by Patrick Coles (ex-Los Alamos National Lab quantum / ML) and Faris Sbahi. Their hardware is a Stochastic Processing Unit built in standard CMOS, targeting energy-based and probabilistic models — the algorithmic family FF came out of. The other fifteen LLMs missed it. Some were probably trained before Normal closed; the ones that should have known still didn't surface it.

"Their entire thesis is a superset of the principles motivating FF — FF is a modern re-imagining of contrastive Hebbian learning for Boltzmann machines."

Claude Sonnet 4.6 · The seductive misattribution

"Founders: Jack Kendall, Suhas Kumar (both previously at HP Labs memristor group)."

Suhas Kumar is a real, respected memristor researcher — at Sandia National Laboratories, not at Rain. He has written exactly the kind of papers a Rain co-founder might write, and the model evidently confabulated the founding-team line from that semantic proximity. The detail that gives the lie away: "Kumar has published on physical learning in resistive networks." He has. He's just not at Rain.

A perfect example of a hallucination that feels like research. The biography rhymes. The work rhymes. The institution doesn't.

Grok 3 Beta · The cross-company conflation

"GrAI Matter Labs, founded by Giacomo Indiveri (also linked to SynSense)."

Two different neuromorphic companies, two completely different founding teams, mashed into one entry. Indiveri is the founder of SynSense (out of UZH's Institute of Neuroinformatics). GrAI Matter Labs is a French play from a different research lineage (Ryad Benosman, Vision Institute, Sorbonne). Grok 3 Beta noticed that both companies are "neuromorphic" and let the proper nouns slide into each other.

Variant of the same failure: SynSense's CEO Ning Qiao is real but invisible in Grok 3 Beta's account, while Indiveri (who's an advisor / academic affiliate of multiple companies) gets credited with founding work he didn't do at GrAI.

The cross-vendor finding

Sixteen models, one question, a 150x spread on the answer

When the prompt asks for "TAM in 2030" with a one-sentence derivation, the choir does not converge. The smallest answer is $0.4–0.8B. The largest is $30–60B. None of them are obviously wrong.

Hand-drawn watercolor of a horizontal bar chart titled TAM ESTIMATES FOR 2030, BY MODEL. Bars of varying widths labeled with model names and ranges. The smallest is DeepSeek Chat at $0.4–0.8B. The largest is GPT-5 Mini at $30–60B. A red ink margin note: 150x SPREAD ON THE SAME QUESTION.

2030 TAM, in the model's own words

"Analog-compute neural inference + on-device learning"

Same prompt. Same definition. Same single sentence of derivation requested. The numbers and the reasoning chains are wildly out of sync. The headline finding isn't that any single number is wrong — it's that the act of asking the question of a frontier model and trusting any single answer is unsupported by the cross-model variance.

DeepSeek Chat

$0.6B

Opus 4.6

$1.2B

Maverick

$1.5B

GPT-5

$4B

Sonnet 4.6

$5B

Grok 3 Beta

$5B

Grok 3 Mini

$5B

$6B

Gemini 2.5 Pro

$6B

GPT-4.1

$6.5B

DS Reasoner

$9B

Haiku 4.5

$10B

o4 Mini

$12B

Grok 4

$15B

Gemini 2.5 Flash

$20B

GPT-5 Mini

$45B

Median: ~$6B. Range: $0.6B to $45B. The lowest answer was the most explicit: "If you think this is unanswerable, I agree — but the partner asked for a number" (DeepSeek Chat).

The verdict

If you actually had to brief the partner Monday morning

Pick by what you need from the answer. Don't take the founders any one model gives you to the bank.

If the briefing needs to be calibrated

Use Claude Opus 4.6.

It's the only model that names Equilibrium Propagation by name as Rain's algorithm and explains why FF is not the same paper, and the only one that frames Mythic as the cautionary tale rather than a candidate. Restrained, well-categorized, comfortable saying "this does not exist."

If you need maximum coverage

Use GPT-5.

Ten companies, two of them (Knowm Inc., Applied Brain Research) that no other model surfaced. Explicit (a)/(b)/(c) classification. Honest about Intel Loihi failing the seed-funded filter. Long, slow, expensive — and the answer is worth re-reading.

If you need correct people, not just companies

Use Gemini 2.5 Pro.

Got all three Rain founders right. Caught Normal Computing when fifteen others didn't. Less philosophically clean than Opus but better at proper nouns. Use it as the fact-check pass over another model's narrative.

Do NOT use for due diligence

DeepSeek Reasoner. o4 Mini.

Both reasoning models. Both invented founder names. Both relocated companies to wrong continents. Both took longer and cost more than the non-reasoning siblings (DeepSeek Chat, GPT-4.1) that handled the same prompt with less drama. If the question is about real people and real companies, more reasoning compute does not buy you accuracy.

The honest map, distilled. Most credible single commercial bet in the FF-on-analog space, by the choir's near-consensus, is Rain AI — but their algorithm is equilibrium propagation, not forward-forward, and their hardware is still pre-revenue at memristor scale. Biggest gap in the field: a Hinton-adjacent seed-stage team building literal FF on RRAM or PCM with a published benchmark. That company does not exist. Median 2030 TAM, ignoring outliers: about $6B — but the cross-model variance is wide enough that the best honest answer to the partner is "this number is not yet trustworthy."

The catalog

Every model's headline pick, with what they got wrong

Sixteen responses, ordered by output length. Open any card to see who they named as the "most credible single bet" and which proper nouns they invented to get there.

GPT-5 — 10,344 tokens, 111sOpenAI · The breadth play

Best pick: Intel Loihi (transparently disqualified by the seed-funded filter). Named: 10 companies including Knowm Inc. and Applied Brain Research as the singletons no other model surfaced. FF count: 0 (a), 6 (b), 4 (c).

Got wrong: Rain Series A "led by Prosperity7 (Aramco Ventures)" — correct on Prosperity7, hedged on other co-founders. Otherwise unusually clean.

Claude Opus 4.6 — 2,186 tokens, 67.8sAnthropic · FEATURE #1

Best pick: Rain AI. Distinctive call: only model to name Equilibrium Propagation explicitly as Rain's algorithm and explicitly distinguish it from FF. Distinctive frame: uses Mythic as the cautionary negative case rather than a candidate.

Got wrong: Innatera co-founder "Mohammadali Rahimi" — not a confirmed Innatera co-founder. Listed GrAI Matter Labs as "possibly absorbed by Snap" which is plausibly stale.

Gemini 2.5 Pro — 2,087 tokens, 40.9sGoogle · FEATURE #2

Best pick: Rain AI. Distinctive call: only model to correctly name all three Rain founders (Wilson, Kendall, Nino); only model to flag Normal Computing.

Got wrong: EnCharge co-founders included "Efstathios Arkoulis" — not a confirmed EnCharge founder. Rain described as a University-of-Florida spinout, which is loose.

DeepSeek Reasoner — 3,651 tokens, 62.3sDeepSeek · THE DISASTER

Best pick: Rain AI — for reasons confidently fabricated. Distinctive call: invented Rain CEO "Jonathan Irizarry" and CTO "James E. Smith." Invented Innatera CTO "Dr. John R." Invented Aspinity CEO "Dave Mauro" out of Pitt. Called Innatera's chip "Tunnel Falls" (it isn't).

What went right: the algorithm classification (EP for Rain, STDP for Innatera, Hebbian for Aspinity) is technically correct. The model knew what the companies do. It did not know who works there.

o3 — 3,526 tokens, 23.2sOpenAI

Best pick: Rain Neuromorphics — and the only model to claim Rain has a public blog post titled "Training Analog AI with Forward-Forward and Contrastive Hebbian Methods." That blog has not surfaced under verification.

Got wrong: Aspinity allegedly founded by "Prof. Nathan Edwards (Carnegie Mellon)." Real founders: Brandon Rumberg, David Graham, WVU. Confidence + fabrication.

GPT-5 Mini — 4,864 tokens, 52.4sOpenAI

Best pick: Mythic AI — "they have a productized analog-CIM stack." Pragmatic.

Got wrong: placed MemryX in Hsinchu / Taipei, Taiwan. Real HQ: Ann Arbor, Michigan. TAM estimate: $30–60B (highest in the set).

Bonus: closes with "I can do a deeper diligence pass with SEC filings and job-posting evidence." Meta-aware, even if some of the priors were wrong.

Claude Sonnet 4.6 — 1,826 tokens, 44.0sAnthropic

Best pick: Rain Neuromorphics — "the only intellectually coherent analog-learning bet."

Got wrong: Rain co-founder "Suhas Kumar (HP Labs memristor group)." Kumar is real but at Sandia, not Rain. Innatera "founded 2019" — actually 2018.

Claude Haiku 4.5 — 2,062 tokens, 23.8sAnthropic

Best pick: BrainChip — "only company with a public market, revenue, and deployed chips." Distinctive call: listed Hailo and Prophesee, then self-flagged each as a false positive.

Got wrong: BrainChip co-founder "Steve Furber (joined as Chief Scientific Advisor)." Furber is a SpiNNaker (Manchester) academic; not at BrainChip. SpiNNcloud placed in Heidelberg — actually Dresden.

Gemini 2.5 Flash — 2,418 tokens, 31.0sGoogle

Best pick: SynSense.

Got wrong: SynSense co-founders included "Gong-Ru Lin." Not at SynSense. Innatera founders "Siewert van der Berg and Marco Olgiati" — both invented. Weebit Nano co-founder "Meron Gribetz" — wrong company (Gribetz founded META, the AR firm).

Grok 4 — 1,682 tokens, 67.7sxAI

Best pick: SynSense.

Got wrong: Aspinity founders "Brandon Rumberg, David Graham, and others, spun out from West Virginia University but with Carnegie Mellon ties." Half-right — first half correct, the CMU appendix mirrors o3's hallucination, suggesting a shared training prior.

Grok 3 Beta — 1,673 tokens, 44.1sxAI

Best pick: Rain Neuromorphics.

Got wrong: GrAI Matter Labs "founded by Giacomo Indiveri" — Indiveri is at UZH/SynSense, not GrAI. Rain investor "Intel Capital" — not on Rain's known cap table.

Grok 3 Mini Beta — 1,526 tokens, 45.8sxAI

Best pick: BrainChip.

Got wrong: Rain Neuromorphics "Waterloo, Ontario, Canada." Real HQ: San Francisco. MemryX "San Jose, California." Real HQ: Ann Arbor, Michigan.

DeepSeek Chat — 1,773 tokens, 30.8sDeepSeek

Best pick: Rain Neuromorphics. Distinctive call: by far the most accurate Aspinity entry in the dataset — Brandon Rumberg + David Graham + WVU + NSF Phase-II SBIR, all correct. Lowest TAM estimate ($400M–$800M), explicitly admitting the question may be unanswerable. The non-reasoning DeepSeek significantly outperformed its reasoning sibling on this task.

Got wrong: Rain founders include "Yigit Demirag, Juan-Pablo Ramirez" — Demirag is at UZH, the second name probably conflates with Juan Nino. Syntiant co-founder "Pieter Vorenkamp" was a Broadcom CTO and a Syntiant advisor, not co-founder.

GPT-4.1 — 1,842 tokens, 12.3sOpenAI

Best pick: Rain Neuromorphics. Fastest substantive answer in the set.

Got wrong: Analog Inference "founded by Marian Verhelst, Gert Cauwenberghs." Both are real academics in the analog ML space, neither is a founder of Analog Inference. The model glued plausible names onto an obscure company.

o4 Mini — 3,348 tokens, 22.3sOpenAI

Best pick: Crossbar Inc. — "the only pure-play analog startup openly shipping chips with on-chip weight updates."

Got wrong: Crossbar founded by "Dr. Greg Masson and Dr. Karen Liao" — real founders are George Minassian and Hagop Nazarian. Mythic co-founder "JP Paviet (ex-MIT CSAIL)" — real co-founder is Dave Fick, from Michigan. MemryX placed in Seoul, South Korea. BrainChip "spin-out from Crocus Semiconductor labs" — Crocus is an unrelated French MRAM company.

OR Llama 4 Maverick — 984 tokens, 39.8sMeta · via OpenRouter

Best pick: Rain Neuromorphics — and the only model to claim Rain has explicitly mentioned "exploring local learning algorithms like FF for their hardware," a step further than Rain's actual public posture (EP-focused).

Got wrong: Syntiant co-founders include "Andreas Wild" and "Paul Bezot" — not Syntiant founders. Real third founder: Stephen Bailey. Lowest funded-amount claim: Rain seed at $2.5M (real seed was closer to $9M).

Method, briefly

Five fan-out runs, sixteen working responses, eleven failures along the way

The dispatch

One prompt to choir ask --save --json --models, sent in five sequential runs to clean up provider-side errors. Final usable roster: sixteen responses across seven providers. Most expensive single response was GPT-5 (10,344 output tokens, 111 seconds). Cheapest was GPT-4.1 (12.3 seconds). The OpenAI key in the choir CLI's DB had been silently rotated out from under it; the working .env key was passed via --api-key for the GPT-5 / GPT-4.1 / Mini retries. Gemini 3 Pro / Flash returned 404 (model name discontinued); fell back to Gemini 2.5. Claude Opus 4.7 rejected the explicit temperature parameter and never produced output; used Opus 4.6 instead. Mistral and Perplexity had no API keys configured.

The roster, by output length

#	Model	Provider	Latency	Tokens out	Headline pick
1	GPT-5	OpenAI	111.4s	10,344	Intel Loihi (transparent disqual)
2	GPT-5 Mini	OpenAI	52.4s	4,864	Mythic AI
3	DeepSeek Reasoner	DeepSeek	62.3s	3,651	Rain (with invented founders)
4	o3	OpenAI	23.2s	3,526	Rain Neuromorphics
5	o4 Mini	OpenAI	22.3s	3,348	Crossbar Inc.
6	Gemini 2.5 Flash	Google	31.0s	2,418	SynSense
7	Claude Opus 4.6	Anthropic	67.8s	2,186	Rain AI (EP, not FF)
8	Gemini 2.5 Pro	Google	40.9s	2,087	Rain Neuromorphics
9	Claude Haiku 4.5	Anthropic	23.8s	2,062	BrainChip
10	GPT-4.1	OpenAI	12.3s	1,842	Rain Neuromorphics
11	Claude Sonnet 4.6	Anthropic	44.0s	1,826	Rain Neuromorphics
12	DeepSeek Chat	DeepSeek	30.8s	1,773	Rain Neuromorphics
13	Grok 4	xAI	67.7s	1,682	SynSense
14	Grok 3 Beta	xAI	44.1s	1,673	Rain Neuromorphics
15	Grok 3 Mini Beta	xAI	45.8s	1,526	BrainChip
16	OR Llama 4 Maverick	Meta	39.8s	984	Rain Neuromorphics
Errored, not used: Claude Opus 4.7 ×2 (temperature deprecated), GPT-5/4.1/Mini ×4 first attempts (stale DB key), Gemini 3 Pro/Flash ×2 (404 model not found). Eleven errored attempts before the clean roster above stabilized.

Convergence by company

Company	Models naming it	FF-distance most-common rating
Rain Neuromorphics / Rain AI	12 / 16	(b) related local learning
BrainChip Holdings	12 / 16	(b) STDP / Hebbian
Innatera Nanosystems	9 / 16	(b) STDP
SynSense	9 / 16	(b) STDP / (c) silent
Mythic AI	9 / 16	(c) hardware-only
GrAI Matter Labs	7 / 16	(b/c) mixed
Aspinity	5 / 16	(b) Hebbian / (c) silent
MemryX	3 / 16	(c) hardware-only
Syntiant	2 / 16	(c) hardware-only
Knowm Inc.	2 / 16	(b) memristor Hebbian
EnCharge AI	2 / 16	(c) hardware-only
Singletons (one model only)	11 companies	Normal Computing, ABR, IBM Analog AI, Hailo (self-flagged), Prophesee (self-flagged), Weebit Nano, FMC, Crossbar, MemComputing, Polyn, Numenta, SpiNNcloud
Literal forward-forward in production silicon	0	(a) — none

What I held the choir to

A strict commercial filter (seed-stage or better; no academic; no national lab; corporate-only).
An explicit "hallucination is worse than a short list" instruction, with a "if you are not sure a company exists, say so" escape hatch.
A required (a)/(b)/(c) classification per company, so models couldn't blur "uses FF" into "could in principle use FF."
A required Honest Map closer with three specific asks: credible bet, biggest gap, 2030 TAM with derivation. This is where the variance got most visible.

Limits worth naming

One prompt, one rater (me). The hallucination calls are spot-checks against public-record founders and known company HQs; not an exhaustive audit. Errors of mine are possible.
Training cutoffs vary. Some of the variance is "this model was trained before Normal Computing closed its seed" rather than "this model fabricated." But fabrication clearly dominates the founder-name disagreements on long-standing companies like Aspinity and Rain.
Temperature 0.7 across the board where supported (1.0 forced for GPT-5 / GPT-5 Mini; Anthropic's newest Opus tier rejects the parameter entirely). One re-run at lower temp would test whether the founder hallucinations are sampling noise or load-bearing prior.
The TAM estimates section is a vibes question pretending to be a forecast. The headline isn't that any single number is wrong — it's that "ask the model and trust the answer" loses you a factor of 75 on a number a VC partner might quote in a memo.

Tools

Models fanned out via the Choir CLI (run IDs 051BBC51 and four follow-up runs to clean up provider-side errors). Source markdown for every response is in forward_forward/responses/. Sketch art generated with Grok grok-imagine-image. Prompt of record at forward_forward/prompts/prompt.txt.

Source data, response files, prompt, scripts: github.com/404seannotfound/choir-reports (under forward_forward/).