Across almost every overlooked scene, the same structure repeats: the pure core is a nonprofit or an academic lab — which is why it's cheap and ignored, and also why you can't just write it a check. The value leaks out through two doors. One: the specific young scientist who spins out (you couldn't invest in the cypherpunk mailing list, but you could have backed Hal Finney). Two: the pick-and-shovel tool everyone in the scene already buys (you couldn't invest in “crypto” in 2010, but you could have sold mining rigs). So the Daxos move is not “pick a scene and deploy” — it's embed cheaply in the two or three best scenes now (sponsorships, relationships, right-of-first-look), back the person when they spin out, and own the shovel where one exists. Only two scenes below have a clean investable pure-play today; the rest are relationships you build before there's a round.
Tier 1 — the strongest scenes
AI interpretability's math underground
Why it ranks here. The purest cypherpunk shape in the whole search: a ~11-person math nonprofit, real shipping code, a coherent fringe worldview, near-zero VC — and it sits on the one bottleneck (can we inspect what we deploy?) that could gate the entire AI industry.
A "training MRI": an LLC-or-successor metric, run during or right after a frontier training run, that flags an emerging capability or a deceptive / sandbagging circuit BEFORE it shows up in behavioral evals — a labeled forward prediction that beats red-team evals and is trusted enough for a lab or regulator to gate a deployment on it. First credible frontier-scale demonstration: ~3-5 years. The "GPT/Bitcoin moment" — a prediction that is adopted as a standard pre-deployment audit — is 5-10 years, and only if forward prediction (not retrodiction) actually lands.
Mostly a WATCH with a relationship option — there is nothing to buy today. Timaeus is a nonprofit: a $100-500K check there is a donation, not equity, and buys no ownership. Harmonic confirms zero SLT commercial vehicle exists. The only investable path is the PEOPLE and the eventual spinout: Daniel Murfet (the mathematician), Jesse Hoogland (ED), Stan van Wingerden (build/eng), Susan Wei. Concrete move: embed now in the devinterp / SLT community (the LessWrong/Alignment-Forum dev-interp threads, the "SLT & Alignment" / Singular Learning Theory summer schools and workshops, the devinterp GitHub), fund a small unrestricted research grant or a compute gift to build goodwill, and secure right-of-first-look on any Murfet/Hoogland/van Wingerden incorporation. The moment they spin a "training-time model-audit / deception-detection" product out, a $100-500K SAFE is the wedge — it buys founder access to the one interpretability angle Goodfire's SAE stack does not own. Secondary watch: Apollo Research — if it converts from nonprofit deception-evals to a commercial audit product, SLT metrics could be its moat, and that entity is closer to a real company.
A relationship plus right-of-first-refusal on the Timaeus core team's eventual for-profit spinout (Hoogland / Murfet / van Wingerden), positioned as the training-time model-audit and deception-detection company — the un-priced interpretability wedge Goodfire's $209M SAE incumbent does not cover. Cost today: a small grant and community presence, not a check.
1) No vehicle — Timaeus stays a nonprofit indefinitely, the team never incorporates, and the check has nowhere to land. 2) Mechanistic interp (Goodfire, SAEs, $209M, Anthropic-backed) wins the entire "interpretability company" market and SLT stays an academic curiosity with no product surface. 3) LLC never scales to frontier models or never produces a real forward prediction — retrodiction forever, so the breakout artifact never arrives. 4) Talent drain: labs (Anthropic, DeepMind) hire the principals into internal safety teams, as they already did with much interp talent, and the independent scene hollows out. 5) The scene is downstream of AI-safety grant money (Open Phil / LTFF); if that funding contracts, the community shrinks before any commercial flywheel forms. 6) Regulatory or lab demand for a "training MRI" never materializes and interp stays an internal cost center.
Bioelectric morphogenetics
Why it ranks here. The single highest-convergence scene — surfaced through 6 of 14 independent lenses. Real, mainstream-respected science with a provocative worldview; the commercial promise runs ahead of the mammalian evidence, so treat it as optionality, not conviction.
The rerating result: a bioelectric intervention — Morphoceuticals' ion-channel drug cocktail or a direct voltage/gap-junction edit — that reproducibly regrows a FUNCTIONAL structure in a MAMMAL in vivo (a mouse or rat digit/limb segment with bone, nerve and vasculature), or normalizes a mammalian tumor in vivo by resetting its bioelectric state. That is the "frog video becomes a mouse video" moment that flips the field from provocative biology to a drug category. Rough timeline: credible mammalian proof-of-concept 2-5 years (Morphoceuticals is aiming squarely here); first-in-human regeneration signal 7-10+ years. A secondary, further-out breakout is Anthrobots as a programmable human-cell delivery/repair platform doing a defined job in vivo. Until a mammalian result replicates, it stays a frog-and-planaria story no matter how many papers publish.
Two concrete entries plus a person to back. (1) Morphoceuticals is the only vehicle actually chasing mammalian translation and has a serious biotech chairman, Roger Pomerantz (ex-Seres CEO, ex-Merck head of licensing), with CEO Jim Jenson — investor contact is literally published ([email protected] / [email protected]). It last raised a $3.5M seed Sept 2024 and is the type to run an A next; a $100-500K check can't lead (Prime Movers Lab anchors) but can take a syndicate slice — what it buys is exposure to the one team trying to make this an actual mammalian drug. (2) The cheaper, better-fit optionality: embed in the Astonishing Labs orbit and back a pre-seed spinout (Anthrobots Inc / Cellja Vu Inc are still ~$0 shells) where $100-500K is genuinely meaningful and you sit next to Jurvetson and Carmack. (3) The person to back is Gizem Gumuskaya — the Anthrobots inventor, now her own PI (MIT/Tufts) and the most natural young founder in the Levin diaspora; a relationship or first check into her spinout is the highest-leverage early move. Watch-list the rest of the diaspora: Douglas Blackiston (xenobots), Patrick McMillen. Honest caveat: this is optionality, not a conviction drug bet — structure it as one small check plus an embed, not a lead.
A single small optionality check into the Astonishing Labs / Gumuskaya orbit at pre-seed — the cheapest, earliest claim on the highest-convergence scientist in the scout, alongside Jurvetson and Carmack — with Morphoceuticals' next round as the alternative syndicate slice for the mammalian-drug shot. Back the founder (Gumuskaya), not just the thesis.
1) Translation may simply fail — planaria/frog regenerative capacity may not transfer to mammals, and 15 years with zero mammalian in-vivo limb/organ regrowth is a real warning, not just early days. 2) The "living robot / software layer" framing is partly hype; a prominent replication failure or a credible "xenobots are just cell clumps" debunk could crater sentiment across the scene. 3) Capital intensity and timeline: a $100-500K seed check gets crushed by dilution over a 10-year drug-development arc — this is not a software moat you compound cheaply. 4) Key-person risk — the whole scene is Levin, who is spread across 350+ papers and multiple vehicles, with IP fragmented across Tufts, Wyss and the studio, so no single company cleanly captures the value. 5) No clear FDA path for voltage-editing therapeutics.
Homebrew & open silicon
Why it ranks here. A proven talent factory whose apex company is already funded by Altman/Friedman/Naval — which validates the thesis while the community that produces the founders stays cheap to embed in.
Two candidate "Bitcoin/GPT moments," ranked by plausibility. Most likely (2-4 yrs, arguably already crossing): the open-EDA + cheap-MPW path becomes the default way a hardware startup gets first custom silicon — a small team takes a genuinely useful custom chip (a real sensor front-end, power-management, or crypto/security block, not a ring-oscillator toy) from Verilog to yielding packaged parts for under $10k using Yosys/OpenROAD + a Tiny-Tapeout-style shuttle. Tiny Tapeout already puts tiny designs on silicon for a few hundred dollars, so this is the artifact closest to breaking out. Less likely but higher-magnitude (5-8 yrs): a bench-top process that repeatably yields useful sub-micron analog/power/MEMS devices at a per-device cost that undercuts a foundry shuttle for low-volume niche parts — a true "desktop fab that makes a part you'd actually ship." The defense variant of this (distributed, air-gapped, trusted-fab for legacy mil nodes) is the one with real dollars behind it and is what Atomic Semi/Fab2 is chasing with DoD money.
Be honest: there is no clean equity round to join today. The apex (Atomic Semi/Fab2) is priced at ~$75M by OpenAI/Friedman/Naval and a $100-500K check almost certainly cannot get in. The genuinely unpriced entities are non-equity: Hackerfab is academic/open-source, Tiny Tapeout is a 4-person service, YosysHQ is a bootstrapped GmbH that has taken zero VC by choice. So the play is TALENT-FIRST, not round-first. Concretely: (1) Embed as the fund-of-record for the Hackerfab diaspora. Sponsor CMU Hackerfab and 1-2 of the six replicating university chapters ($25-50k of the check as sponsorship/equipment grants buys founder relationships and first-look), and build a direct line to the named principals — founders Elio Bourcart and Alexander Hakim, advisor Tathagata Srimani (CMU faculty, the credible bridge to a spinout), and Matt Venn (Tiny Tapeout) as the hub of everyone taping out. (2) The $100-500K first check goes into the NEXT Atomic-Semi-shaped spinout out of this orbit BEFORE Friedman/OpenAI price it — a Hackerfab-founder team productizing either a bench-fab tool or an open-EDA-to-silicon SaaS ("Vercel for tapeout"). (3) Watch YosysHQ / whitequark (Amaranth) for the day a commercial open-EDA cloud company forms around that stack — that is the most investable adjacent shape and is currently unpriced. Net: mostly WATCH-AND-EMBED now, with one cheap sponsorship wedge that converts to first-check rights on the spinout.
Back the person, not the movement: get first-check rights on the CMU Hackerfab / Zeloof-orbit founder who spins out next, via a cheap sponsorship embed today. The scene reliably manufactures unusually real hardware talent (measured-yield devices, six-university replication) and its last apex founder got funded by OpenAI and Nat Friedman — Daxos wants to be the fund those founders already know before the brand-name seed prices them.
(1) Benchtop physics ceiling — 10um / ~1,000 transistors has no large commercial market; the "RepRap for chips" analogy is false because the output isn't a shippable end-use part. (2) Willingness-to-pay disproven — Efabless went bankrupt; the open community loves free tools and won't pay, so value accrues to the mature-node foundries (SkyWater, GlobalFoundries) not the open layer. (3) The scene's winner already left "open" for proprietary defense tooling (Atomic Semi/Fab2, DoD/USAF money), so the commercial gravity is reshoring capex, not open silicon — a game Daxos can't fund at $100-500K. (4) No investable vehicle: unpriced entities are non-profit/bootstrapped, priced entity is unreachable, so the thesis depends entirely on a future spinout that may go straight to the OpenAI/AI-Grant crowd and never need Daxos. (5) Long, capital-heavy timelines for anything past the tooling/EDA layer.
Active inference
Why it ranks here. The one scene here with a clean, US, investable pure-play today (Noumenal Labs) that stripped the AGI story and aimed the theory at a near-term robotics-safety wedge.
A pure or hybrid active-inference world-model agent that decisively out-samples deep/model-based RL on a recognized hard benchmark - Atari-100k SOTA, or better, a real robot manipulation/locomotion skill learned in ~10x fewer trials with calibrated uncertainty driving safe exploration - published AND independently reproduced. The industrial variant: an RxInfer-style real-time Bayesian controller/anomaly detector deployed on edge hardware with hard ROI versus a deep net. Rough timeline: 3-7 years, high variance. Most probable form of the "win" is NOT a pure-AIF banner but AIF ideas (calibrated uncertainty, active/information-seeking exploration, structured world-models) absorbed into a robotics foundation-model stack - which would validate the worldview while stranding pure-play companies.
The investable pure-play is Noumenal Labs (US, NYC; founded 2024 by ex-VERSES/FEP researchers, robotics autonomy with "calibrated uncertainty"). Real deep-tech seed VCs are already in (Cyber Fund, Mila Ventures, Machina, Capital Factory), so a $100-500K check most likely rides a seed-extension or Series A insider allocation; the relationship buys diligence access, technical reference into the Friston diaspora, and co-invest rights - fits Daxos's US preference. Cheapest true-first-check: Lazy Dynamics (RxInfer / TU Eindhoven, Bert de Vries + Dmitry Bagaev) - $0 institutional, 8 people, a genuinely fast working artifact; $100-500K could be a real priced first check with meaningful ownership, but it is EU and the wedge (real-time Bayesian inference for control/edge) is niche. Do NOT buy VERSES stock - public, collapsing, a value trap dressed as contrarianism. Community embed: sponsor/participate in the Active Inference Institute (Daniel Friedman) and live in the pymdp / RxInfer GitHub communities to spot the next diaspora founder who pairs AIF with a concrete robotics/control benchmark BEFORE they raise. Verify Noumenal's exact cap table and founder identities directly (Harmonic hasn't captured the round) before any check.
Back the team that already stripped the AGI story and pointed active inference at a near-term robotics-safety wedge - Noumenal Labs (calibrated-uncertainty autonomy) as the lead target, with Lazy Dynamics as cheap EU optionality - not the theory, the institute, or the public vehicle.
(1) No capability edge: calibrated uncertainty and sample efficiency are delivered just as well by deep ensembles / Bayesian deep learning without the FEP baggage, so AIF adds philosophy, not performance. (2) The decisive benchmark win never arrives and AIF stays a describe-anything framework. (3) Ideas and talent get absorbed into big-lab world models (JEPA, Dreamer successors), leaving no standalone company to own. (4) Reputational contagion: VERSES's very public collapse taints the whole "active inference" label with generalist investors for years. (5) The honest near-term wedges (RxInfer control/edge inference) are real but sub-venture-scale niches.
Altermagnetism & spin-physics
Why it ranks here. The cleanest 'sell shovels to the overlooked' play in the set — you don't have to pick which exotic magnetism wins, you sell the measurement tool every lab already buys.
A working altermagnetic memory cell: room-temperature, non-volatile, with electrical WRITE (current- or ~0.1V-gate-driven Neel-vector switching) AND electrical READ (anomalous-Hall or tunneling magnetoresistance) in a CMOS-compatible thin film at sub-femtojoule/bit, published and then independently reproduced, beating STT/SOT-MRAM on the energy-speed-density frontier. That is the Bitcoin/GPT moment: the first altermagnetic MRAM bit that is manifestly better than incumbent MRAM at room temp. The CISS analog would be a magnet-free chiral-molecule spin injector integrated into a solid-state junction with reproducible, mechanism-explained room-temp polarization. Rough timeline: a convincing single lab-demo cell in 3-6 years (AMSwitch targets exactly this), a manufacturable/integrable substrate 8-12 years out.
Be honest: the altermagnet layer itself is NOT investable today, it is grant-funded academia (AMSwitch/Chalmers, Mainz, Nottingham, Czech Academy) with no equity for a $100-500K check to buy. Three real moves, in order of fit. (1) PICK-AND-SHOVEL, the only thing actually investable now: every altermagnet and CISS lab images spin domains with NV-diamond scanning magnetometers. QZabre LLC (ETH Zurich spinout, 22 staff, bootstrapped, $0 institutional VC, sells to exactly these labs, an AMSwitch partner) is the standout, a check or relationship could open its first priced round; Qnami is the alternative cap table. This monetizes today and rides whichever spin substrate wins, altermagnet or CISS or neither. TAM is small (academic + industrial R&D), so it is a tool-co outcome, not a rocket. (2) WATCH + EMBED at the US anchor: Igor Mazin at George Mason University is the American altermagnet theory center, build a right-of-first-look relationship with his group and with any MnTe/CrSb thin-film growth group at MIT/Cornell/Berkeley, so Daxos sees the first device spinout before anyone. Cost is time, not capital. (3) FOUND-WITH-A-PROFESSOR, highest upside, hardest: seed a US thin-film/materials-IP or tool co with a group member as technical cofounder, but this needs a deep-tech syndicate and more than a $100-500K seed. Net: one narrow pick-and-shovel entry (QZabre) plus a watch-and-embed posture at Mazin/GMU; the physics is a watch, not a buy.
Back the measurement-tool layer, concretely QZabre (ETH spinout, bootstrapped, $0 VC, already selling NV-diamond magnetometers to altermagnet labs), as the pick-and-shovel that earns revenue today and rides every spin-physics substrate, while building a right-of-first-look relationship with Igor Mazin's group at George Mason, the US altermagnet anchor, for the eventual device spinout.
1. Altermagnetic memory may never beat incumbent SOT/STT-MRAM or SRAM on real power-performance-area and could stay a lab curiosity for 20 years like multiferroics. 2. CISS mechanism is unresolved, part of the effect could be measurement artifact, device reproducibility may never arrive. 3. Horizon of 8-12 years to a manufacturable substrate is far longer than a $100-500K seed's patience, and there is no equity vehicle at the physics layer to hold anyway. 4. The confirmed materials (MnTe, CrSb) may resist CMOS integration; the RuO2 retraction shows headline materials can evaporate. 5. The only investable vehicle, scanning NV magnetometry, is a small niche market (low hundreds of $M TAM), QZabre is bootstrapped and may not want VC, and the space is capital-heavy hardware in Europe, outside Daxos's US software-speed comfort zone.
Tier 2 — real and overlooked, but watch or hard to enter
Computing that isn't GPUs
Why it ranks here. The energy wall of AI is real and this attacks it, but the startup edge (Extropic, Normal) is already funded and the unpriced part is still in the lab.
A room-temperature, actually-fabricated chip - p-bit, dopant-network, spintronic, or coupled-learning silicon - that an independent third party benchmarks as beating a shipping digital NPU (e.g. a mobile-SoC NPU) on energy-per-inference for a real edge workload (keyword spotting, always-on anomaly detection, sensor fusion), AND that ships Entry: Mostly a WATCH-and-relationship scene with one or two targeted seed shots, not a place to deploy the full thesis. A $100-500K check buys nothing meaningful in Extropic's or Normal's next round and shouldn't chase them. Where it does buy something: (a) first-check into a fresh spinout at formation from the unpriced academic
Malleable software
Why it ranks here. A beautiful worldview with real artifacts, but the core is nonprofits/consultancies and the local-first sibling is already taking venture money.
An LLM-native successor to the spreadsheet / HyperCard: a mainstream "moldable app" where a non-programmer describes and then keeps editing a real personal tool, not a chat transcript but a persistent app they own, that runs offline, syncs across their devices, and can be forked and re-molded by the next person. The Bitcoin/GPT Entry: Most of the pure core is NOT investable: Ink & Switch, Hazel, Folk/Dynamicland, and feenk are lab/academic/consultancy structures that will not take a venture check, and Automerge is MIT-licensed so there is no equity in the artifact itself. The right move is to back the founder turning a real malleable artifact into a shipping LLM-in
Connectomics & brain preservation
Why it ranks here. The least-VC-touched scene of all and genuinely important, but structurally a nonprofit/FRO world on a 2030s timeline — embed, don't check.
Two-stage. Near milestone (structural): a complete whole mouse-brain connectome (~1000x the fly, ~exabyte scale) under NIH's BRAIN CONNECTS program — plausibly early-to-mid 2030s. The true "GPT/Bitcoin moment" (functional): a reconstructed small-nervous-system connectome — C. elegans or a fly circuit — that is simulated and Entry: Mostly WATCH + one relationship, not a fund-defining check. The core is structurally near-uninvestable: FROs (E11, Forest) and Astera are grant-funded nonprofits with no equity to buy. Three concrete moves. (1) Embed in the Convergent Research FRO + Astera ecosystem now — FROs are explicitly designed to open-source or SPIN OUT a for-profi
Living bioelectronics
Why it ranks here. Surprising, real, near-zero VC — but a single-lab, atoms-not-bits bet years from a manufacturable material. One node worth a relationship (Malvankar, Yale).
A fermentation-manufacturable conductive protein: grams-to-kilograms of an engineered cytochrome nanowire (OmcZ/OmcS-class) or a de-novo conductive protein produced in E. coli, cast into a film or electrode with published, independently-reproduced metal-approaching conductivity plus a self-healing or biodegradable property no metal offers Entry: The single investable node is Nikhil Malvankar (Yale Microbial Sciences Institute). He owns the correct mechanism (cytochrome nanowires, not pili), the highest-conductivity biological wires, and the engineered/recombinant work closest to a manufacturable material — he is the one who would spin out "grown conductive protein / self-hea
Condensed-matter nuclear (LENR)
Why it ranks here. The reverse-signal you asked for — actively dismissed, quietly government-funded — but the highest crank risk in the set. A relationship and a pre-committed check on a clean replication, not a dollar today.
One independently replicated, third-party-instrumented, sustained excess-heat result — an anomalous power output cleanly above measurement error, reproduced across at least two labs, accompanied by a coherent nuclear signature (transmutation products or commensurate radiation) — coming out of the ARPA-E capability-team measurements at a n Entry: Watch-and-relationship, NOT a check today. The investable surface is empty for a US micro-VC: Brillouin is a US Series-C zombie (avoid), Clean Planet is Japanese and corporate/government-controlled (hard to enter, wrong geography), and every other entity is a forum, a blog, or a one-person LLC. A $100-500K check buys nothing here except o
Programmable cryptography
Why it ranks here. Real and cypherpunk to the core, but the honest verdict is it's already priced: a16z-crypto and Paradigm own the infra layer. Only a narrow US app-layer entry remains.
Two candidates, and the likely winner is NOT the FHE one. (1) Most probable "GPT moment," 2-4 years: client-side ZK proofs of identity/provenance/authenticity become a default web primitive — proving you are a unique human, that a document or media asset is genuine, or a fact about your bank/credentials without revealing them — Entry: Do NOT chase the infra layer — it is priced and owned by a16z crypto, Paradigm, Multicoin, Blockchain Capital, Protocol Labs, Founders Fund, and Daxos (small US equity first-checks, non-crypto-native) has zero edge there. The only real entry has two moves. (1) Embed to source: the unpriced asset is the applied-cryptographer talent pool gr
Tier 3 — scout watchlist (surfaced, not yet deep-verified)
- Ambient-pressure superconductor recovery. The diamond-anvil underground already makes materials superconduct near room temperature; the only unsolved problem is getting them out of the anvil. Zero VC, pure physics.
- Metal fuels (iron-powder energy). Burn iron powder for carbon-free heat, reduce the rust back with green hydrogen — a Dutch-Canadian combustion scene, near-zero VC.
- Mechanochemistry. Room-temperature reactions in ball mills that normally need extreme heat/pressure/solvent — including solvent-free ammonia. Green-chemistry alpha, barely funded.
- Molecular programming / DNA computing. The survivors of the Drexler nanotech bust: write a program, matter folds and computes. FNANO scene, mostly unpriced.
- Thorium-229 nuclear clock. In 2024 physicists excited an atomic nucleus with a tabletop laser — a clock based on the nucleus, ~10x better than atomic clocks. No venture money.
- Masakhane / African NLP. A 1,000+ person 'for Africans, by Africans' community that built data + models for 40+ languages Western labs ignored. Barely funded.
- Ukraine's Brave1 robotic-warfare scene. Thousands of engineers shipping attritable autonomous drones/robots on days-long loops under real conditions — partly circled by US defense VC, the room itself still under-mapped.
- E-graphs / equality saturation. A small PL-research band rebuilt an old compiler trick into a general optimization engine; quietly spreading into ML compilers, EDA, and theorem proving. No VC.
- Applied category theory (Topos). Making 'how systems compose' executable and provable — the longest-horizon, least-certain bet, essentially no VC.
1. Open a relationship with the Timaeus / Singular Learning Theory people now (Daniel Murfet, Jesse Hoogland) and get right-of-first-look on the for-profit spinout — this is the closest thing to “the cypherpunks, today,” and there is nothing to buy yet, which is exactly the point. 2. Back a person out of the bioelectric or homebrew-silicon scenes — Gizem Gumuskaya (the Anthrobots inventor, spinning out near Jurvetson and Carmack) or the next founder out of CMU Hackerfab — with one small optionality check each. 3. Take the one clean shovel: QZabre's NV-diamond magnetometers already sell into every altermagnet and CISS lab with zero VC on the cap table — revenue today, and it rides whichever spin-physics horse wins.