hraness

side quests

the pursuits around the company — the founders' earlier work, the competition and the internship, and the neo-lab ecosystem.

the founders' earlier work

  1. 2015–2018the quant years · endedtworek's pre-openai career — senior quantitative researcher at independent view in amsterdam, building fully systematic futures signals as the fund's most senior researcher. the trading frame survives: 'being a good trader means understanding your algorithm very well… in the end all you see is the p&l.'linkedin, 2026-09-16warsaw.ai, 2020blog.agihouse.org, 2026-07-29
  2. 2026–million integrals · continuingmillionintegrals.com — the self-authored public record of the openai work: copilot, codex, humaneval, plugins and code interpreter, o1, o3, gpt-5, the chatgpt agent. part cv, part deposition.millionintegrals.com, 2026-09-16
  3. 2018–the shampoo→psgd line · continuingthe optimizer line anil carries in: shampoo, co-created at google brain with vineet gupta and tomer koren, then the psgd framework that 'can recover shampoo… a much better framework' — implemented inside the internal coreauto codebase. the qr-factorization kernel competition (~60× over cusolver, per anil on training data) sits in the same lineage.the-information-bottleneck.com, 2026-09-10digg, 2026teahose.com, 2026-07-29scholar.google.com, 2026-09-16
  4. 2024–2026project mariner · handed-offthe agents-and-computer-use line — anmol gulati as tech lead on project mariner in gemini, kanav garg on the same project ('one of the most stimulating 0->1 projects'). the multimodal-agent experience the lab's work-automation thesis needs.linkedin, 2026-09-16linkedin, 2026-09-16scholar.google.com, 2026-09-16
  5. 2021–2026model behavior and openai labs · handed-offjang's openai arc — product lead on dall·e 2, gpt-4, the chatgpt api, tts, and embeddings; founding lead of model behavior and steward of the model spec; then openai labs, prototyping interfaces 'beyond chat or even agents,' reportedly adjacent to the jony ive effort.linkedin, 2026-09-16prometheusroot.com, 2026womenpm.org, undated
  6. 2025–reservoir samples · continuingreservoir samples — jang's own essays on model personality and human-ai relationships, running alongside the lab work.prometheusroot.com, 2026joannejang.com, 2026-09-16
  7. 2025–the agi-storytelling project · continuingthe stated reason villagra left openai — 'art, music and storytelling to help people understand the transition to artificial general intelligence.' how it coexists with an executive-officer seat at a research lab is unreported.business insider, 2025-08linkedin, 2026-09-16
  8. 2021–gpu mode · continuingsaroufim's other company — the discord community for gpu kernel engineering that became a business running public kernel competitions, plus 'popcorn,' training an llm to write kernels. the qr-factorization contest anil cites lives in this lineage. saroufim remains a pytorch maintainer.marksaroufim.com, 2026-09snipd, 2026linkedin, 2026-09-16teahose.com, 2026-07-29
  9. 2020–2024mosaicml composer and adept · handed-offlamp's earlier stack — early engineer at mosaicml, where his resume credits the composer training framework and the yahp hyperparameter system; mts at adept ai labs; a stealth cofounder in between; google deepmind after.averylamp.me, 2026-09-16linkedin, 2026-09-16
  10. 2025–the fysk seat · continuinglamp's community seat — 'longtime engineer and lp at fysk,' the founders you should know community whose showcase hosted tworek on september 16.newsletter.foundersysk.com, 2026-08-25

the company's quests

  1. 2026-05 –the blog corpus · continuingthe company blog — five posts in 2026: saroufim on boring authoring and ai-written systems code, the one layer deeper launch and results, amid on neural architecture discovery. the lab's public instrumentation.coreauto.com, 2026-05-27
  2. 2026-08 – 2026-09one layer deeper · endedthe public research competition co-run with tilde research: repeated modular squaring — learn the computation, then run longer chains than seen in training. hard track: one h100, one hour of training, certify depth t=1…64. results september 9: 15,602 submissions, only 18 hard uploads from 5 accounts certified. the first public artifact of the research program.coreauto.com, 2026-05-27linkedin, 2026-09-16mcleish7.github.io, 2026-09-16
  3. 2026-09 – 2027-03the agm program · activethe associate general manager program — four interns in sf, january–march 2027, $10k/month plus profit share, each running real online businesses on the company's agent infrastructure with a budget. 'by watching where people intervene, we want to understand what agents still need to learn.' a hiring instrument and a data-generation scheme in one.coreauto.com, 2026-09-16
  4. 2026-07the pacing letter · endedthe 'pacing the frontier' open letter — 1,386 frontier-ai employees requesting 'technical and governance tools needed to deliberately pace the frontier of automated ai development.' core automation employees are among the signatories per the hn discussion — the automation lab signing the pacing letter.pacingthefrontier.com, 2026-07hacker news, 2026-07-28

the ecosystem

  1. 2026–tilde research · continuingthe collaborator lab — a ~12-person sf 'moonshot' interpretability/architectures outfit co-founded by dhruv pai and mason wang; co-ran one layer deeper and published 'wall attention' mit-licensed in june 2026.linkedin, 2026-09-16aiinsiders.net, 2026-06-02mcleish7.github.io, 2026-09-16
  2. 2026-08the alumni ventures feeder · ended'avsf – core automation 2026, llc' — the alumni ventures feeder spv with its own form d: $2,109,250 sold to 100 investors, first sale august 3. the only investor provable by filing.sec, 2026-08-17
  3. 2025–the neo-lab cohort · continuingthe cohort core automation gets grouped with — thinking machines (murati), ssi (sutskever), periodic labs, ami labs (lecun), adaption labs (hooker), flapping airplanes ($180m, data-efficiency), inherent, prime intellect — competing for the same frontier-talent pool and the same neolab capital.the-decoder.com, 2026-04-22ai2.work, 2026-04-22the-decoder.com, 2026-01-29hacker news, 2026-07-28

AI-drafted at Ben Guo's direct request and credited to Hraness; every claim links to its cataloged source.