Aemon growth playbook

Build a proof page at /results/circle-packing: the exact variant (number of circles, unit square, sum of radii), Aemon's value next to the 2025 value and today's best public value, the coordinates as a downloadable file, a short checker anyone can run in a notebook (non-overlap, inside the square, sum of radii), the run's compute cost and wall-clock time, and a short log of how the session got there. Point both "View Aemon's solution" and "Verify it yourself" at it, put the number itself in the homepage section, and submit the result to the public record table where it qualifies. If later results beat it, keep a dated history on the page.

Aemon sells to R&D teams, and researchers trust numbers they can rerun. A checkable result can be shared, cited and posted to technical communities, where an unverifiable one gets questioned. It also separates Aemon from rivals making similar claims, because few of them publish everything.

[ ] Exact variant: n circles, unit square, sum of radii
[ ] Aemon's value, the 2025 value and today's best public value
[ ] Coordinates as a downloadable file
[ ] Checker script and one-click notebook
[ ] Compute cost and wall-clock time
[ ] Short run log: hypotheses tried, what worked
[ ] Both homepage links point here
[ ] Number shown on the homepage
[ ] Submitted to the public record table if it qualifies

• Publishing the number invites comparison with newer records

  Show today's best public value honestly and date the result; the cost and method are the story.

• Someone finds an error in the checker

  Use exact, well-known checks and invite corrections publicly.

• Your websiteHomepage (checked November 10, 2026): "A world-record result, found by Aemon": one session surpassed "Google DeepMind's 2025 AlphaEvolve result" on Circle Packing, "a classic NP-hard optimization problem", for "< $10" of compute. The page states no score and no circle count; the only link in the section, "Verify it yourself", opens DeepMind's own alphaevolve_results notebook, and "View Aemon's solution" has no link target in the page as served.Source
• Competitor researchCognizant AI Lab (August 6, 2026): its agent found a 27-circle packing with sum of radii 2.685978684198, beating the record of 2.685350025228 that had stood since 2011, then beat 21 live Packomania records across n = 2 to 93; the solver, an independent checker and every result are published on GitHub.See the research Source
• Competitor researchShinkaEvolve: "an open-source framework (Apache 2.0 License)" for evolutionary code optimization with LLMs; it reports a new state-of-the-art 26-circle packing found with only 150 samples, which its authors say outperforms AlphaEvolve's.See the research Source
• Y CombinatorCompany profile: Winter 2026 batch, founded 2025, team of 3, San Francisco. Aemon reads state-of-the-art research, then generates and evolves many candidate approaches against the user's evaluation criteria, and says it set a new world record on circle packing for under $10 of compute. The founders are twin brothers: the CEO is an international medalist in math and robotics; the CTO has published LLM research at ICLR and EMNLP.Source

Split it in two. Make the circle-packing replay show packing terms: the eval (non-overlap, bounds, sum of radii), the value climbing past the old record and the final layout. Then, if Aemon has improved a search or ranking pipeline, publish that as its own short case with the before and after Recall@10, anonymised if needed.

Consistency builds trust with technical readers, and a retrieval case speaks to the many teams that run search and ranking in production.

• No shareable retrieval customer

  Run it on a public retrieval benchmark instead.

• Your websiteHomepage replay and hero: the replay diagram is labelled "Eval specification", "BM25", "Reranker", "top-k" and "Recall@10 baseline 0.000", retrieval terms beside a circle-packing caption, and is captioned "Replay visualization — not the product UI, and not the limit of what Aemon can solve." The hero counters "Research progress" and "SOTA Found" read "0%" in the page as served.Source
• Your websiteHomepage offer: H1 "Your next breakthrough, found by AI."; "Backed by Y Combinator"; "Talk to us" and "Book a Demo" buttons, with a cal.com booking embed; "From first call to technical breakthroughs in two weeks" in four steps (Eval Setup, Evaluation Locked, Aemon Deployed, Live & Running), where Aemon connects to the customer's codebase and research context; the steps appear twice. No pricing, customer logos, case studies, named use cases or industries. "Aemon is an applied research team focused on building the future of R&D."Source

Add four questions to the booking: the problem in one sentence, the metric that defines better, the current baseline, and whether a scoring script exists today. Route eval-ready teams to a technical founder, and send everyone else a one-page eval template that says what to bring. Use the answers to prepare the call.

Qualifying on the thing the product needs most shortens the path to the first result and protects scarce founder time.

• Fewer bookings

  Keep three fields optional and watch booking volume.

• Your websiteHomepage offer: H1 "Your next breakthrough, found by AI."; "Backed by Y Combinator"; "Talk to us" and "Book a Demo" buttons, with a cal.com booking embed; "From first call to technical breakthroughs in two weeks" in four steps (Eval Setup, Evaluation Locked, Aemon Deployed, Live & Running), where Aemon connects to the customer's codebase and research context; the steps appear twice. No pricing, customer logos, case studies, named use cases or industries. "Aemon is an applied research team focused on building the future of R&D."Source
• ResearchPositioning history: the headline in February 2026 was "The AI Research Engineer" and is now "Your next breakthrough, found by AI."; the page title is "Aemon | The AI R&D Engineer"; the meta description is "Give Aemon a problem and an eval. It discovers optimal solutions, often going beyond what human experts can find."; the Y Combinator profile says "The Forward-Deployed AI Research Engineer". The domain was registered in November 2025; no Hacker News mentions and no news mentions in the last 30 days were found.Source

Publish a privacy policy and terms, plus a plain-language /security page: what access Aemon needs, where code and data are processed and stored, who owns discovered improvements and code, retention and deletion, whether model providers can train on customer data, and the option of running inside the customer's environment. Link it from the footer and the booking confirmation.

Removing the first security question shortens procurement for enterprise R&D teams, and IP ownership terms answer the question every buyer has about discoveries.

• Promising more than the team can support

  Describe only what is true today and mark what is planned.

• Your websiteSitemap and site (November 10, 2026): the sitemap lists one URL, the homepage, last updated September 4, 2026; /privacy returns 404 and the page links to no privacy policy or terms; no blog, pricing page, docs or changelog; LinkedIn is the only social profile linked; the product sits on app and accounts subdomains.Source
• Your websiteHomepage offer: H1 "Your next breakthrough, found by AI."; "Backed by Y Combinator"; "Talk to us" and "Book a Demo" buttons, with a cal.com booking embed; "From first call to technical breakthroughs in two weeks" in four steps (Eval Setup, Evaluation Locked, Aemon Deployed, Live & Running), where Aemon connects to the customer's codebase and research context; the steps appear twice. No pricing, customer logos, case studies, named use cases or industries. "Aemon is an applied research team focused on building the future of R&D."Source
• Competitor researchAlphaEvolve on Google Cloud (July 9, 2026): "generally available to all Google Cloud customers on Gemini Enterprise Agent Platform" after private preview since December; users supply a baseline algorithm and goals and get optimized, readable code; named users BASF, JetBrains and Kinaxis; examples include chip design, logistics routing and medical research.See the research Source

Pick one category phrase. Forward-deployed is the real difference, because an Aemon engineer locks the eval. Use it in the H1 or subhead, the title, the meta description, the YC profile, LinkedIn and job posts. Keep the breakthrough line as the promise beneath it.

Consistent naming builds recall and helps searchers and assistants file Aemon in the right category.

• Picking a phrase nobody searches

  Keep R&D and research engineer as words in the description.

• ResearchPositioning history: the headline in February 2026 was "The AI Research Engineer" and is now "Your next breakthrough, found by AI."; the page title is "Aemon | The AI R&D Engineer"; the meta description is "Give Aemon a problem and an eval. It discovers optimal solutions, often going beyond what human experts can find."; the Y Combinator profile says "The Forward-Deployed AI Research Engineer". The domain was registered in November 2025; no Hacker News mentions and no news mentions in the last 30 days were found.Source
• Y CombinatorCompany profile: Winter 2026 batch, founded 2025, team of 3, San Francisco. Aemon reads state-of-the-art research, then generates and evolves many candidate approaches against the user's evaluation criteria, and says it set a new world record on circle packing for under $10 of compute. The founders are twin brothers: the CEO is an international medalist in math and robotics; the CTO has published LLM research at ICLR and EMNLP.Source

Compress or stream the replay media, serve images in modern formats at display size, set long cache lifetimes on static assets, and fix the one console error. Aim for under 3 MB on first load.

A lighter page loads fast for the people a champion forwards it to, on any connection, and is cheaper to serve.

• A blurrier replay

  Compare quality on a phone before shipping.

• PageSpeedMobile score 99, desktop 100, accessibility 100, SEO 100 (tested November 10, 2026). The homepage downloads 31,793 KB in 25 requests; suggested savings are 8.1 MB from cache lifetimes, 573 KB from image delivery and 0.3 s of render-blocking requests. Real visitors (Chrome UX Report) see LCP of 1.8 s and CLS of 0.01 at the 75th percentile, so the measured Core Web Vitals pass; one browser console error is logged.Source

Render the final values in the HTML (for example the record value, the number of results beaten and the compute cost) and animate up to them, rather than starting the HTML at zero. Add the same facts to the meta description.

Machine readers increasingly introduce companies to buyers; giving them the real numbers costs nothing.

• Numbers go stale

  Pull them from the results page in idea 2.

• Your websiteHomepage replay and hero: the replay diagram is labelled "Eval specification", "BM25", "Reranker", "top-k" and "Recall@10 baseline 0.000", retrieval terms beside a circle-packing caption, and is captioned "Replay visualization — not the product UI, and not the limit of what Aemon can solve." The hero counters "Research progress" and "SOTA Found" read "0%" in the page as served.Source
• ResearchPositioning history: the headline in February 2026 was "The AI Research Engineer" and is now "Your next breakthrough, found by AI."; the page title is "Aemon | The AI R&D Engineer"; the meta description is "Give Aemon a problem and an eval. It discovers optimal solutions, often going beyond what human experts can find."; the Y Combinator profile says "The Forward-Deployed AI Research Engineer". The domain was registered in November 2025; no Hacker News mentions and no news mentions in the last 30 days were found.Source

Run Aemon across a full range of circle counts and two or three other problems with public best-known tables, using the checker from idea 1. Publish a results page that lists, for each instance, the public best value, Aemon's value, the cost and a download, marking wins, ties and losses honestly. Update it as new runs finish.

Breadth of verified results is the strongest evidence that a search system generalises, and every new win is a reason to post again.

• Losses on many instances

  Publish them; honesty is the point, and losses show where to improve.

• #1 Make the world-record claim checkable: publish Aemon's score, coordinates and a checker instead of linking to the old record

• Competitor researchCognizant AI Lab (August 6, 2026): its agent found a 27-circle packing with sum of radii 2.685978684198, beating the record of 2.685350025228 that had stood since 2011, then beat 21 live Packomania records across n = 2 to 93; the solver, an independent checker and every result are published on GitHub.See the research Source
• Your websiteHomepage (checked November 10, 2026): "A world-record result, found by Aemon": one session surpassed "Google DeepMind's 2025 AlphaEvolve result" on Circle Packing, "a classic NP-hard optimization problem", for "< $10" of compute. The page states no score and no circle count; the only link in the section, "Verify it yourself", opens DeepMind's own alphaevolve_results notebook, and "View Aemon's solution" has no link target in the page as served.Source
• HiringMember of Technical Staff (Research Engineering): $180K–$250K, San Francisco or remote, visa sponsorship, starting with a four-month paid work trial at $15,000 a month. Work areas: long-horizon agents, post-training and RL, hypothesis generation and search, codebase understanding, experiment orchestration, benchmark and evaluation design, research infrastructure and product surfaces for autonomous research systems. Backed by tier-one investors and angels including Paul Graham; no customers listed.Source

Write five pages, one for each problem type where a scoring function usually already exists: GPU and CPU kernels (latency), compilers and schedulers (runtime), search and ranking (recall, NDCG), routing and scheduling (cost), and placement or layout (area, wirelength). Each page covers what the eval looks like, what Aemon would try, what a two-week pilot delivers, and an eval template to download.

Buyers search by their problem, not by a category name, and each page pre-qualifies teams who already have a metric.

• Spreading too thin

  Start with two pages and add as demand shows.

• #4 Ask "do you have an eval?" before the demo, and send a free eval template to everyone who books

• Your websiteSitemap and site (November 10, 2026): the sitemap lists one URL, the homepage, last updated September 4, 2026; /privacy returns 404 and the page links to no privacy policy or terms; no blog, pricing page, docs or changelog; LinkedIn is the only social profile linked; the product sits on app and accounts subdomains.Source
• Your websiteHomepage offer: H1 "Your next breakthrough, found by AI."; "Backed by Y Combinator"; "Talk to us" and "Book a Demo" buttons, with a cal.com booking embed; "From first call to technical breakthroughs in two weeks" in four steps (Eval Setup, Evaluation Locked, Aemon Deployed, Live & Running), where Aemon connects to the customer's codebase and research context; the steps appear twice. No pricing, customer logos, case studies, named use cases or industries. "Aemon is an applied research team focused on building the future of R&D."Source
• Competitor researchAlphaEvolve on Google Cloud (July 9, 2026): "generally available to all Google Cloud customers on Gemini Enterprise Agent Platform" after private preview since December; users supply a baseline algorithm and goals and get optimized, readable code; named users BASF, JetBrains and Kinaxis; examples include chip design, logistics routing and medical research.See the research Source
• Your websiteHomepage replay and hero: the replay diagram is labelled "Eval specification", "BM25", "Reranker", "top-k" and "Recall@10 baseline 0.000", retrieval terms beside a circle-packing caption, and is captioned "Replay visualization — not the product UI, and not the limit of what Aemon can solve." The hero counters "Research progress" and "SOTA Found" read "0%" in the page as served.Source

Package the four steps as an Eval Sprint: a fixed fee, a locked eval agreed on day three, two weeks of Aemon runs, and a written report of every improvement with code. Add a simple term: if Aemon doesn't beat the locked baseline, the customer pays only for the eval work. Publish a short page with what is included, the timeline and a starting price or range.

A fixed price and a clear risk-sharing term turn an open-ended consulting conversation into a decision a team lead can sign, and the locked eval makes the guarantee measurable.

• Arguments about whether the baseline was beaten

  The locked eval decides, agreed in writing before the runs start.

• #4 Ask "do you have an eval?" before the demo, and send a free eval template to everyone who books

• Your websiteHomepage offer: H1 "Your next breakthrough, found by AI."; "Backed by Y Combinator"; "Talk to us" and "Book a Demo" buttons, with a cal.com booking embed; "From first call to technical breakthroughs in two weeks" in four steps (Eval Setup, Evaluation Locked, Aemon Deployed, Live & Running), where Aemon connects to the customer's codebase and research context; the steps appear twice. No pricing, customer logos, case studies, named use cases or industries. "Aemon is an applied research team focused on building the future of R&D."Source
• Your websiteSitemap and site (November 10, 2026): the sitemap lists one URL, the homepage, last updated September 4, 2026; /privacy returns 404 and the page links to no privacy policy or terms; no blog, pricing page, docs or changelog; LinkedIn is the only social profile linked; the product sits on app and accounts subdomains.Source

Pick ten widely used libraries with clear benchmark suites (numerics, parsing, compression, search). Run Aemon against each suite under a fixed budget, keep only improvements that pass the project's tests, and open well-documented pull requests saying how they were found. Track merged PRs on an Aemon page and write up the best ones.

Merged upstream changes are third-party verification, reach the exact engineers who buy, and show Aemon working in real codebases, not puzzles.

• Maintainers see AI PRs as spam

  Follow each project's rules, send few, high-quality PRs, and review every change by hand.

• Competitor researchWeco AI: "We build recursively self-improving AI"; a public weco-cli on GitHub and docs; AIDE reached "Human-level performance on data science competitions" and its agent is described as the top contributor in a public model-training challenge.See the research Source
• Your websiteHomepage (checked November 10, 2026): "A world-record result, found by Aemon": one session surpassed "Google DeepMind's 2025 AlphaEvolve result" on Circle Packing, "a classic NP-hard optimization problem", for "< $10" of compute. The page states no score and no circle count; the only link in the section, "Verify it yourself", opens DeepMind's own alphaevolve_results notebook, and "View Aemon's solution" has no link target in the page as served.Source
• HiringMember of Technical Staff (Research Engineering): $180K–$250K, San Francisco or remote, visa sponsorship, starting with a four-month paid work trial at $15,000 a month. Work areas: long-horizon agents, post-training and RL, hypothesis generation and search, codebase understanding, experiment orchestration, benchmark and evaluation design, research infrastructure and product surfaces for autonomous research systems. Backed by tier-one investors and angels including Paul Graham; no customers listed.Source

Add /log with a short, dated post for each notable run: the problem, the eval, what Aemon tried, what failed and the final result with a checker. Launch the first post (the record from idea 1) on Hacker News and in relevant research communities, with the founders answering questions. Aim for two posts a month.

Technical audiences reward transparent, checkable work, and a steady log builds search presence and credibility for a new domain.

• A launch that gets picked apart

  Launch only with the checker and full numbers from idea 1.

• #1 Make the world-record claim checkable: publish Aemon's score, coordinates and a checker instead of linking to the old record

• Your websiteSitemap and site (November 10, 2026): the sitemap lists one URL, the homepage, last updated September 4, 2026; /privacy returns 404 and the page links to no privacy policy or terms; no blog, pricing page, docs or changelog; LinkedIn is the only social profile linked; the product sits on app and accounts subdomains.Source
• ResearchPositioning history: the headline in February 2026 was "The AI Research Engineer" and is now "Your next breakthrough, found by AI."; the page title is "Aemon | The AI R&D Engineer"; the meta description is "Give Aemon a problem and an eval. It discovers optimal solutions, often going beyond what human experts can find."; the Y Combinator profile says "The Forward-Deployed AI Research Engineer". The domain was registered in November 2025; no Hacker News mentions and no news mentions in the last 30 days were found.Source
• Y CombinatorCompany profile: Winter 2026 batch, founded 2025, team of 3, San Francisco. Aemon reads state-of-the-art research, then generates and evolves many candidate approaches against the user's evaluation criteria, and says it set a new world record on circle packing for under $10 of compute. The founders are twin brothers: the CEO is an international medalist in math and robotics; the CTO has published LLM research at ICLR and EMNLP.Source

Publish one page comparing three routes: do it yourself with an open-source framework, use a cloud agent, or use Aemon forward-deployed. Compare them on who builds the eval, time to first result, what runs in your codebase, cost, and who is accountable for the result. Be fair about when DIY is the better choice.

Technical buyers will make this comparison anyway; owning it puts Aemon's real advantage, the eval work and accountability, in front of them.

• Claims go out of date

  Date the page and cite each source.

• Competitor researchShinkaEvolve: "an open-source framework (Apache 2.0 License)" for evolutionary code optimization with LLMs; it reports a new state-of-the-art 26-circle packing found with only 150 samples, which its authors say outperforms AlphaEvolve's.See the research Source
• Competitor researchAlphaEvolve on Google Cloud (July 9, 2026): "generally available to all Google Cloud customers on Gemini Enterprise Agent Platform" after private preview since December; users supply a baseline algorithm and goals and get optimized, readable code; named users BASF, JetBrains and Kinaxis; examples include chip design, logistics routing and medical research.See the research Source
• Competitor researchWeco AI: "We build recursively self-improving AI"; a public weco-cli on GitHub and docs; AIDE reached "Human-level performance on data science competitions" and its agent is described as the top contributor in a public model-training challenge.See the research Source
• Your websiteHomepage offer: H1 "Your next breakthrough, found by AI."; "Backed by Y Combinator"; "Talk to us" and "Book a Demo" buttons, with a cal.com booking embed; "From first call to technical breakthroughs in two weeks" in four steps (Eval Setup, Evaluation Locked, Aemon Deployed, Live & Running), where Aemon connects to the customer's codebase and research context; the steps appear twice. No pricing, customer logos, case studies, named use cases or industries. "Aemon is an applied research team focused on building the future of R&D."Source

Generate a one-page weekly report from the run logs: the metric against the locked baseline with a chart, the three most interesting hypotheses tried (including failures), the improvements shipped as code, and the next experiments. End with one suggested next problem.

Visible progress keeps the champion engaged and gives them material to share, which drives renewals and expansion.

• Weeks with no gains

  Report what was ruled out; that is progress too.

• Your websiteHomepage offer: H1 "Your next breakthrough, found by AI."; "Backed by Y Combinator"; "Talk to us" and "Book a Demo" buttons, with a cal.com booking embed; "From first call to technical breakthroughs in two weeks" in four steps (Eval Setup, Evaluation Locked, Aemon Deployed, Live & Running), where Aemon connects to the customer's codebase and research context; the steps appear twice. No pricing, customer logos, case studies, named use cases or industries. "Aemon is an applied research team focused on building the future of R&D."Source
• HiringMember of Technical Staff (Research Engineering): $180K–$250K, San Francisco or remote, visa sponsorship, starting with a four-month paid work trial at $15,000 a month. Work areas: long-horizon agents, post-training and RL, hypothesis generation and search, codebase understanding, experiment orchestration, benchmark and evaluation design, research infrastructure and product surfaces for autonomous research systems. Backed by tier-one investors and angels including Paul Graham; no customers listed.Source

At the end of each successful sprint, ask the champion for one or two internal introductions, give them a short internal one-pager with their own results, and offer the second team a shorter setup, because security and contracts are already done. Track internal referrals as their own pipeline.

Internal referrals reuse trust, procurement and security approval, so they close faster than new logos.

• Asking too early

  Only after a result beats the baseline.

• #14 Send every customer a weekly breakthrough report: gains against the locked baseline, what was tried, what's next

• Your websiteHomepage offer: H1 "Your next breakthrough, found by AI."; "Backed by Y Combinator"; "Talk to us" and "Book a Demo" buttons, with a cal.com booking embed; "From first call to technical breakthroughs in two weeks" in four steps (Eval Setup, Evaluation Locked, Aemon Deployed, Live & Running), where Aemon connects to the customer's codebase and research context; the steps appear twice. No pricing, customer logos, case studies, named use cases or industries. "Aemon is an applied research team focused on building the future of R&D."Source
• Y CombinatorCompany profile: Winter 2026 batch, founded 2025, team of 3, San Francisco. Aemon reads state-of-the-art research, then generates and evolves many candidate approaches against the user's evaluation criteria, and says it set a new world record on circle packing for under $10 of compute. The founders are twin brothers: the CEO is an international medalist in math and robotics; the CTO has published LLM research at ICLR and EMNLP.Source

Post a limited YC offer: five discounted Eval Sprints for companies with a scoring script ready, in exchange for permission to publish the result and the company's name. Prioritise companies whose problem matches the use-case pages, and turn each result into a case study.

Batchmates are the fastest route to named, public proof, and each case study makes the next enterprise sale easier.

• Small startups have weak evals

  Accept only teams with a working scoring script.

• #10 Sell the two-week promise as a fixed-price sprint: beat the locked baseline, or the second week is free

• Y CombinatorCompany profile: Winter 2026 batch, founded 2025, team of 3, San Francisco. Aemon reads state-of-the-art research, then generates and evolves many candidate approaches against the user's evaluation criteria, and says it set a new world record on circle packing for under $10 of compute. The founders are twin brothers: the CEO is an international medalist in math and robotics; the CTO has published LLM research at ICLR and EMNLP.Source
• Your websiteHomepage offer: H1 "Your next breakthrough, found by AI."; "Backed by Y Combinator"; "Talk to us" and "Book a Demo" buttons, with a cal.com booking embed; "From first call to technical breakthroughs in two weeks" in four steps (Eval Setup, Evaluation Locked, Aemon Deployed, Live & Running), where Aemon connects to the customer's codebase and research context; the steps appear twice. No pricing, customer logos, case studies, named use cases or industries. "Aemon is an applied research team focused on building the future of R&D."Source
• HiringMember of Technical Staff (Research Engineering): $180K–$250K, San Francisco or remote, visa sponsorship, starting with a four-month paid work trial at $15,000 a month. Work areas: long-horizon agents, post-training and RL, hypothesis generation and search, codebase understanding, experiment orchestration, benchmark and evaluation design, research infrastructure and product surfaces for autonomous research systems. Backed by tier-one investors and angels including Paul Graham; no customers listed.Source

Release a minimal open-source package: a YAML eval spec, a runner that scores a candidate against it, regression checks, and three examples (circle packing, a ranking task, a kernel benchmark). Link it from the site, the eval template and the booking flow. Teams who build a spec with it can attach it when they book.

It turns Aemon's hardest-to-scale step into something buyers do themselves, and every public repo using it is a qualified lead.

• Giving away the method

  Open the eval format, keep the search system closed.

• #4 Ask "do you have an eval?" before the demo, and send a free eval template to everyone who books

• HiringMember of Technical Staff (Research Engineering): $180K–$250K, San Francisco or remote, visa sponsorship, starting with a four-month paid work trial at $15,000 a month. Work areas: long-horizon agents, post-training and RL, hypothesis generation and search, codebase understanding, experiment orchestration, benchmark and evaluation design, research infrastructure and product surfaces for autonomous research systems. Backed by tier-one investors and angels including Paul Graham; no customers listed.Source
• ResearchPositioning history: the headline in February 2026 was "The AI Research Engineer" and is now "Your next breakthrough, found by AI."; the page title is "Aemon | The AI R&D Engineer"; the meta description is "Give Aemon a problem and an eval. It discovers optimal solutions, often going beyond what human experts can find."; the Y Combinator profile says "The Forward-Deployed AI Research Engineer". The domain was registered in November 2025; no Hacker News mentions and no news mentions in the last 30 days were found.Source
• Competitor researchWeco AI: "We build recursively self-improving AI"; a public weco-cli on GitHub and docs; AIDE reached "Human-level performance on data science competitions" and its agent is described as the top contributor in a public model-training challenge.See the research Source

For the Live & Running phase, test a fee tied to the measured improvement: a base retainer plus a share of the value of gains on the locked metric (for example compute saved per month or latency reduced), capped per quarter. Keep the fixed-price sprint as the entry point. Agree the formula when the eval is locked.

It aligns price with the customer's value and makes the buying decision easier for a finance team: they pay more only when Aemon delivers more.

• Gains are hard to value in money

  Offer it only where the metric maps to cost, such as compute or time.

• #10 Sell the two-week promise as a fixed-price sprint: beat the locked baseline, or the second week is free
• #14 Send every customer a weekly breakthrough report: gains against the locked baseline, what was tried, what's next

• Your websiteHomepage (checked November 10, 2026): "A world-record result, found by Aemon": one session surpassed "Google DeepMind's 2025 AlphaEvolve result" on Circle Packing, "a classic NP-hard optimization problem", for "< $10" of compute. The page states no score and no circle count; the only link in the section, "Verify it yourself", opens DeepMind's own alphaevolve_results notebook, and "View Aemon's solution" has no link target in the page as served.Source
• Your websiteHomepage offer: H1 "Your next breakthrough, found by AI."; "Backed by Y Combinator"; "Talk to us" and "Book a Demo" buttons, with a cal.com booking embed; "From first call to technical breakthroughs in two weeks" in four steps (Eval Setup, Evaluation Locked, Aemon Deployed, Live & Running), where Aemon connects to the customer's codebase and research context; the steps appear twice. No pricing, customer logos, case studies, named use cases or industries. "Aemon is an applied research team focused on building the future of R&D."Source
• Your websiteSitemap and site (November 10, 2026): the sitemap lists one URL, the homepage, last updated September 4, 2026; /privacy returns 404 and the page links to no privacy policy or terms; no blog, pricing page, docs or changelog; LinkedIn is the only social profile linked; the product sits on app and accounts subdomains.Source

Open a waitlisted self-serve tier: upload an eval spec and a repository, set a compute budget, and get results and code with the same reports as managed customers. Keep forward-deployed engagements for hard evals and enterprises. Start with problem types where evals are simple and safe to run in a sandbox.

Self-serve lets demand grow faster than the team and brings in the developers who later champion bigger deals.

• Bad evals give bad results and churn

  Require the spec checker to pass before a run starts.

• #17 Open-source the eval spec format and checker harness, so teams arrive with an Aemon-ready eval

• Your websiteSitemap and site (November 10, 2026): the sitemap lists one URL, the homepage, last updated September 4, 2026; /privacy returns 404 and the page links to no privacy policy or terms; no blog, pricing page, docs or changelog; LinkedIn is the only social profile linked; the product sits on app and accounts subdomains.Source
• Competitor researchWeco AI: "We build recursively self-improving AI"; a public weco-cli on GitHub and docs; AIDE reached "Human-level performance on data science competitions" and its agent is described as the top contributor in a public model-training challenge.See the research Source
• Competitor researchShinkaEvolve: "an open-source framework (Apache 2.0 License)" for evolutionary code optimization with LLMs; it reports a new state-of-the-art 26-circle packing found with only 150 samples, which its authors say outperforms AlphaEvolve's.See the research Source
• Competitor researchAlphaEvolve on Google Cloud (July 9, 2026): "generally available to all Google Cloud customers on Gemini Enterprise Agent Platform" after private preview since December; users supply a baseline algorithm and goals and get optimized, readable code; named users BASF, JetBrains and Kinaxis; examples include chip design, logistics routing and medical research.See the research Source

Publish a list of ten well-known open optimisation and construction problems with public best-known values. Run Aemon on them continuously, publish every improvement with the construction and a checker, and ask independent mathematicians to confirm major results before announcing them. Report progress quarterly.

Verified results on famous problems are the most credible demo an R&D agent can have, and they attract both customers and the research talent Aemon is hiring.

• No new results for months

  Report close misses and techniques; pick a mix of hard and tractable problems.

• #1 Make the world-record claim checkable: publish Aemon's score, coordinates and a checker instead of linking to the old record
• #2 A public record ladder: run Aemon across every circle count and post each result against the public tables

• Your websiteHomepage (checked November 10, 2026): "A world-record result, found by Aemon": one session surpassed "Google DeepMind's 2025 AlphaEvolve result" on Circle Packing, "a classic NP-hard optimization problem", for "< $10" of compute. The page states no score and no circle count; the only link in the section, "Verify it yourself", opens DeepMind's own alphaevolve_results notebook, and "View Aemon's solution" has no link target in the page as served.Source
• Competitor researchCognizant AI Lab (August 6, 2026): its agent found a 27-circle packing with sum of radii 2.685978684198, beating the record of 2.685350025228 that had stood since 2011, then beat 21 live Packomania records across n = 2 to 93; the solver, an independent checker and every result are published on GitHub.See the research Source
• HiringMember of Technical Staff Intern 2027: a four-month internship in San Francisco at $8K–$15K, with Python, TypeScript, machine learning and AI agents; exceptional interns may be considered for a founding technical role; olympiad and competition results are valued, and the interview includes an open-ended problem to define, structure and solve.Source

Use the first sprints and the use-case pages to choose the industry with the clearest evals, highest value per gain and fastest buyers. Build a vertical page, a library of eval templates for its typical problems and a named case study, then run targeted outreach to the twenty companies most like the reference.

Buyers in one industry trust peers, and repeated evals in one domain make every next sprint faster and better.

• Choosing the wrong industry

  Decide on evidence from real calls, and review after two quarters.

• #9 Use-case pages for problems that already have an eval: kernels, compilers, search ranking, scheduling and routing
• #16 Offer a YC-only sprint to deep-tech batchmates with a metric to beat, in exchange for a public case study

• Your websiteHomepage offer: H1 "Your next breakthrough, found by AI."; "Backed by Y Combinator"; "Talk to us" and "Book a Demo" buttons, with a cal.com booking embed; "From first call to technical breakthroughs in two weeks" in four steps (Eval Setup, Evaluation Locked, Aemon Deployed, Live & Running), where Aemon connects to the customer's codebase and research context; the steps appear twice. No pricing, customer logos, case studies, named use cases or industries. "Aemon is an applied research team focused on building the future of R&D."Source
• Competitor researchAlphaEvolve on Google Cloud (July 9, 2026): "generally available to all Google Cloud customers on Gemini Enterprise Agent Platform" after private preview since December; users supply a baseline algorithm and goals and get optimized, readable code; named users BASF, JetBrains and Kinaxis; examples include chip design, logistics routing and medical research.See the research Source

Publish one hard, well-scored problem each quarter with an eval and checker. Show a leaderboard with Aemon's score as a marked entry. Invite strong human submissions to a fast-track founder interview, and write up how the best humans and Aemon differed.

It reaches researchers through competition, which they enjoy, gives an honest, public comparison of humans and Aemon, and fills the hiring funnel at the same time.

• Humans beat Aemon

  Publish it anyway; it is honest and makes a better story.

• #17 Open-source the eval spec format and checker harness, so teams arrive with an Aemon-ready eval

• HiringMember of Technical Staff (Research Engineering): $180K–$250K, San Francisco or remote, visa sponsorship, starting with a four-month paid work trial at $15,000 a month. Work areas: long-horizon agents, post-training and RL, hypothesis generation and search, codebase understanding, experiment orchestration, benchmark and evaluation design, research infrastructure and product surfaces for autonomous research systems. Backed by tier-one investors and angels including Paul Graham; no customers listed.Source
• HiringMember of Technical Staff Intern 2027: a four-month internship in San Francisco at $8K–$15K, with Python, TypeScript, machine learning and AI agents; exceptional interns may be considered for a founding technical role; olympiad and competition results are valued, and the interview includes an open-ended problem to define, structure and solve.Source

Add compute cost and wall-clock time to every published result (proof page, record ladder, open-source PRs, case studies with permission), and show a running median cost per verified improvement on the homepage. Explain how it's measured.

A concrete, low cost per result reframes buying Aemon as an obvious experiment and gives press and posts a simple number to repeat.

• Buyers expect every problem to cost $10

  Show the range and explain that hard evals cost more; the engineer's time is priced separately.

• #1 Make the world-record claim checkable: publish Aemon's score, coordinates and a checker instead of linking to the old record

• Your websiteHomepage (checked November 10, 2026): "A world-record result, found by Aemon": one session surpassed "Google DeepMind's 2025 AlphaEvolve result" on Circle Packing, "a classic NP-hard optimization problem", for "< $10" of compute. The page states no score and no circle count; the only link in the section, "Verify it yourself", opens DeepMind's own alphaevolve_results notebook, and "View Aemon's solution" has no link target in the page as served.Source
• Y CombinatorCompany profile: Winter 2026 batch, founded 2025, team of 3, San Francisco. Aemon reads state-of-the-art research, then generates and evolves many candidate approaches against the user's evaluation criteria, and says it set a new world record on circle packing for under $10 of compute. The founders are twin brothers: the CEO is an international medalist in math and robotics; the CTO has published LLM research at ICLR and EMNLP.Source