Credit
The Independent 100 list was curated by dan (@irl_danB) and published on 2026-09-24 at independent.prose.md (announcement). The ranks and section names are his, and number 100 is his open wild-card slot. This study is ours and is not endorsed by him or by anyone on the list.
What they publicly do
PhD student at MIT CSAIL in the OASYS lab, co-advised by Tim Kraska and Omar Khattab [Verified omarkhattab.com]; first author of Recursive Language Models (NeurIPS 2026) and author of Speculative Programmatic Tool Calling (Aug 2026) [Verified alexzhang13.github.io/blog/2025/rlm, github.com/alexzhang13/spec-ptc]. He promotes the GPU MODE kernel competitions [Reported: X profile x.com/a1zhang].
- X profile: x.com/a1zhang
- GitHub: github.com/alexzhang13
- Site: alexzhang13.github.io
- Blog: alexzhang13.github.io/blog/2025/rlm
- Also: github.com/gpu-mode
Public work
- rlm repo MIT last activity 2026-09-24 · 5,646 GitHub stars on 2026-09-24
Inference library for RLMs (pip install rlms): the prompt is stored as a REPL variable and the root LM peeks, greps, partitions and recursively sub-queries it; supports several sandboxes and includes a verifiers training environment [Verified README].
- Recursive Language Models (paper) paper no license stated last activity 2026
RLM(GPT-5-mini) more than doubled GPT-5's correct answers on an OOLONG split at similar cost, and held up at 10M+ tokens on BrowseComp-Plus subsets [Verified blog; results not reproduced].
- spec-ptc repo MIT last activity 2026-08-24 · 214 GitHub stars on 2026-09-24
Speculative programmatic tool calling: while the model streams REPL code, completed statements launch pure tool calls early in a shadow REPL; side-effecting tools are never speculated [Verified README].
- rlm-minimal repo MIT last activity 2026-01-07 · 862 GitHub stars on 2026-09-24
Gist-sized RLM implementation with a REPL environment [Verified repo].
- VideoGameBench repo license file present, not recognised by GitHub last activity 2025-05-30 · 373 GitHub stars on 2026-09-24
Benchmark environment for vision-language models playing video games [Verified repo].
- GPU MODE reference-kernels (promoted, not authored) repo license file present, not recognised by GitHub last activity 2026-09-18 · 310 GitHub stars on 2026-09-24
Problem sets for the GPU MODE kernel leaderboard, which he promotes [Verified repo; his role is Inference].
Licenses of GitHub repositories are from the authenticated GitHub license API on 2026-09-25, one row per repository in evidence/licenses.jsonl (“no license file” means the API found none; “license file present, not recognised by GitHub” means GitHub could not classify it). Other licenses are as the source states. Check the license file before reusing anything.
Relevance to our work
5 of 5 lenses: agents, memory, orchestration, evals, research-method
RLMs treat long context as data the model inspects with code and sub-calls, the problem long agent sessions run into; rlm and spec-ptc are MIT and small [Verified: GitHub license API, 2026-09-25, evidence/licenses.jsonl].
Deep dive:
- Omar Khattab and alex zhang, GEPA for skills and Recursive Language Models for long context (written 2026-09-25)
Scored 0 to 5 against the study’s relevance rubric: 5 means public, openly licensed work squarely in our lanes that we could study or adopt now, and 1 means little connection to our work. The score is one reviewer’s judgement of how much this public work bears on our own work. It is not a ranking of people and says nothing about the person or the quality of their work.
What we would study next
- rlm-minimal (github.com/alexzhang13/rlm-minimal). The smallest reference implementation of the method, in an isolated environment only (it runs model-written code in-process). Effort: small.
- spec-ptc blog (alexzhang13.github.io/blog/2026/spec-ptc). Its rule that only tools marked pure may run speculatively. Effort: small.
- RLM paper (arxiv.org/abs/2512.24601). Full method, ablations and cost behavior before any adoption decision. Effort: medium.
Before reusing anything: rlm, rlm-minimal and spec-ptc are MIT. rlm's default REPL runs model-written code via exec in the host process, so use one of its sandbox backends. spec-ptc's daemon defaults to a socket in the system temp directory; set a different socket path where that matters. GPU MODE reference-kernels and kernelbot use a non-OSI 'June 9 Researcher Reciprocity License' (RAIL-style use restrictions). videogamebench LICENSE is MIT text but GitHub reports NOASSERTION.
Sources
- alexzhang13.github.io/blog/2025/rlm
- github.com/alexzhang13/rlm
- github.com/alexzhang13/rlm-minimal
- github.com/alexzhang13/spec-ptc
- arxiv.org/abs/2512.24601
- github.com/alexzhang13/videogamebench
- github.com/gpu-mode/reference-kernels
- omarkhattab.com
- x.com/a1zhang/status/2091938825580716079
Overall confidence in this record: High. Sources were read on 2026-09-24 and every link was checked on 2026-09-25. The full record is line 20 of people.jsonl.
Limits
Not verified: His role in GPU MODE beyond promotion. The 2026 'mgh' blog post (alexzhang13.github.io/blog/2026/mgh) was not read. RLM benchmark numbers not reproduced.
This record comes from a first-pass survey of public material by an AI agent session on 2026-09-24, edited for publication on 2026-09-25. Nothing was installed or run for it. Roles, stars and licenses change; see the study’s limits.
Request a correction
Is something about this record wrong or out of date, or would you rather not be covered? Open a correction (a GitHub issue form with this page’s address filled in). A request to remove a record is honoured without argument; accepted corrections are recorded in the changelog with the date.