We are a team of independent researchers from Germany working on ARC AGI 2 since last summer. The general opinion on open-weight models is that they are too weak for this fairly difficult benchmark and score at near noise levels. We found that GPT OSS 120B is actually much more capable than previously thought, once the interleaved thinking regime is stabilized. We basically let the model use a stateful IPython based REPL via function calling and patched vLLM so that the model can reliably do interleaved thinking. The score jumped more than 4x.
Technical write-up: https://pivotools.github.io/posts/agentic_coding_arc_agi/\nCode: https://github.com/gutfeeling/arc-agi-2-submission\nData: https://huggingface.co/datasets/arcagi2/arcagi2-agentic-codi...
For safety, we support sandboxed execution using IPyBox (local Docker) and Daytona (cloud), so others can reproduce this without running untrusted code locally.
It gets more interesting: the effect seems to be general and translates seamlessly to other models without even changing prompts. We are not sure why agentic coding is so powerful in ARC AGI 2, which isn't traditionally thought of as an agentic benchmark. Perhaps code execution provides a stronger form of verification than COT, or perhaps it encourages a qualitatively different form of thinking.
We will be around for a while and would be happy to hear ideas / feedback and discuss infra issues / interleaved thinking / GPT OSS / ARC AGI 2."},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: We let GPT OSS 120B write and run Python and ARC AGI 2 jumped 4x"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/gutfeeling/arc-agi-2-submission"}},"_tags":["story","author_steinsgate","story_46972709","show_hn"],"author":"steinsgate","created_at":"2026-02-11T09:16:19Z","created_at_i":1770801379,"num_comments":0,"objectID":"46972709","points":2,"story_id":46972709,"story_text":"Hi HN,
We are a team of independent researchers from Germany working on ARC AGI 2 since last summer. The general opinion on open-weight models is that they are too weak for this fairly difficult benchmark and score at near noise levels. We found that GPT OSS 120B is actually much more capable than previously thought, once the interleaved thinking regime is stabilized. We basically let the model use a stateful IPython based REPL via function calling and patched vLLM so that the model can reliably do interleaved thinking. The score jumped more than 4x.
Technical write-up: https://pivotools.github.io/posts/agentic_coding_arc_agi/\nCode: https://github.com/gutfeeling/arc-agi-2-submission\nData: https://huggingface.co/datasets/arcagi2/arcagi2-agentic-codi...
For safety, we support sandboxed execution using IPyBox (local Docker) and Daytona (cloud), so others can reproduce this without running untrusted code locally.
It gets more interesting: the effect seems to be general and translates seamlessly to other models without even changing prompts. We are not sure why agentic coding is so powerful in ARC AGI 2, which isn't traditionally thought of as an agentic benchmark. Perhaps code execution provides a stronger form of verification than COT, or perhaps it encourages a qualitatively different form of thinking.
We will be around for a while and would be happy to hear ideas / feedback and discuss infra issues / interleaved thinking / GPT OSS / ARC AGI 2.","title":"Show HN: We let GPT OSS 120B write and run Python and ARC AGI 2 jumped 4x","updated_at":"2026-03-05T23:34:35Z","url":"https://github.com/gutfeeling/arc-agi-2-submission"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"GeorgeHaldane"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["picojool"],"value":"Hello HN.
There is no shortage of excellent JSON libraries out there, so initially this one was built mostly out of curiosity for writing parsers.
JSON proved to be a quite pleasant format to work with so eventually this project evolved into a proper feature complete lib with an API quite similar to nlohmann_json, but much better performance (~2x-5x depending on the workload) and a very concise codebase (single header ~1k L.O.C., slightly larger than picojson), in short:
- Concise single-header JSON lib\n - All JSON types map to std:: containers\n - Performance better than nlohmann_json / picojson, some metrics even beat RapidJSON\n - Supports reflection\n - Supports conversions for custom containers\nhttps://github.com/DmitriBogdanov/UTL/blob/master/docs/modul..."},"title":{"matchLevel":"none","matchedWords":[],"value":"Show HN: Yet another JSON lib (C++17)"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/DmitriBogdanov/UTL/blob/master/docs/module_json.md"}},"_tags":["story","author_GeorgeHaldane","story_44453564","show_hn"],"author":"GeorgeHaldane","created_at":"2025-07-03T10:31:21Z","created_at_i":1751538681,"num_comments":0,"objectID":"44453564","points":2,"story_id":44453564,"story_text":"Hello HN.
There is no shortage of excellent JSON libraries out there, so initially this one was built mostly out of curiosity for writing parsers.
JSON proved to be a quite pleasant format to work with so eventually this project evolved into a proper feature complete lib with an API quite similar to nlohmann_json, but much better performance (~2x-5x depending on the workload) and a very concise codebase (single header ~1k L.O.C., slightly larger than picojson), in short:
- Concise single-header JSON lib\n - All JSON types map to std:: containers\n - Performance better than nlohmann_json / picojson, some metrics even beat RapidJSON\n - Supports reflection\n - Supports conversions for custom containers\nhttps://github.com/DmitriBogdanov/UTL/blob/master/docs/modul...","title":"Show HN: Yet another JSON lib (C++17)","updated_at":"2025-07-03T11:02:55Z","url":"https://github.com/DmitriBogdanov/UTL/blob/master/docs/module_json.md"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"jv22222"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["picojool"],"value":"--
I could not find a single link that nicely summed up the effort happening at Distributed Bio / Centivax - With that in mind, here are a few different pages and videos about this interesting project.
--
Announcement about successfully engineering potent picomolar antibodies that block known neutralizing ACE2 epitopes, preventing the novel coronavirus from infecting human cells.
https://www.youtube.com/watch?v=BzX2qlyn_MM
A video about how they are using ML to analyze and mutate their existing CoV1 antibodies to block CoV2.
https://www.youtube.com/watch?v=LkbycGC_fsQ
A Centivax page (not sure how it fits with Distributed Bio but seems to be connected) that explains the project as a whole.
https://www.centivax.com/covid19
A page that describes the underlying effort to engineer monoclonal therapeutics against 21st-century pathogens through computationally guided antibody discovery.
https://www.centivax.com/monoclonals
This all sounds very hopeful and interesting to me. That said, it would be great to hear from some folks on HN who have a deeper understanding of all of this."},"title":{"matchLevel":"none","matchedWords":[],"value":"Evolving a cure for Covid-19 using computationally guided antibody discovery"}},"_tags":["story","author_jv22222","story_22810756","ask_hn"],"author":"jv22222","created_at":"2020-04-08T07:23:33Z","created_at_i":1586330613,"num_comments":0,"objectID":"22810756","points":1,"story_id":22810756,"story_text":"--
I could not find a single link that nicely summed up the effort happening at Distributed Bio / Centivax - With that in mind, here are a few different pages and videos about this interesting project.
--
Announcement about successfully engineering potent picomolar antibodies that block known neutralizing ACE2 epitopes, preventing the novel coronavirus from infecting human cells.
https://www.youtube.com/watch?v=BzX2qlyn_MM
A video about how they are using ML to analyze and mutate their existing CoV1 antibodies to block CoV2.
https://www.youtube.com/watch?v=LkbycGC_fsQ
A Centivax page (not sure how it fits with Distributed Bio but seems to be connected) that explains the project as a whole.
https://www.centivax.com/covid19
A page that describes the underlying effort to engineer monoclonal therapeutics against 21st-century pathogens through computationally guided antibody discovery.
https://www.centivax.com/monoclonals
This all sounds very hopeful and interesting to me. That said, it would be great to hear from some folks on HN who have a deeper understanding of all of this.","title":"Evolving a cure for Covid-19 using computationally guided antibody discovery","updated_at":"2024-09-20T05:58:28Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"steinsgate"},"title":{"matchLevel":"none","matchedWords":[],"value":"Agentic coding improves ARC AGI 2 performance across models"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["picojool"],"value":"https://pivotools.github.io/pivotools-quarto-blog/posts/agentic_coding_arc_agi/"}},"_tags":["story","author_steinsgate","story_46945095"],"author":"steinsgate","children":[46945096],"created_at":"2026-02-09T13:42:32Z","created_at_i":1770644552,"num_comments":1,"objectID":"46945095","points":1,"story_id":46945095,"title":"Agentic coding improves ARC AGI 2 performance across models","updated_at":"2026-03-05T23:33:07Z","url":"https://pivotools.github.io/pivotools-quarto-blog/posts/agentic_coding_arc_agi/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"jmkim"},"title":{"matchLevel":"none","matchedWords":[],"value":"Apple Pay breaks record with 1M subscribers in South Korea on first day"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["picojool"],"value":"https://pickool.net/apple-pay-breaks-record-with-1-million-subscribers-in-s-korea-on-first-day/"}},"_tags":["story","author_jmkim","story_35257163"],"author":"jmkim","children":[35257527],"created_at":"2023-03-22T05:25:10Z","created_at_i":1679462710,"num_comments":1,"objectID":"35257163","points":1,"story_id":35257163,"title":"Apple Pay breaks record with 1M subscribers in South Korea on first day","updated_at":"2024-09-20T13:39:17Z","url":"https://pickool.net/apple-pay-breaks-record-with-1-million-subscribers-in-s-korea-on-first-day/"},{"_highlightResult":{"author":{"fullyHighlighted":true,"matchLevel":"full","matchedWords":["picojool"],"value":"picololo"},"story_text":{"matchLevel":"none","matchedWords":[],"value":"Fnuz : A dedicated search engine to free online movies, animes, cartoons and tv-shows."},"title":{"matchLevel":"none","matchedWords":[],"value":"Free streaming searchengine"},"url":{"matchLevel":"none","matchedWords":[],"value":"http://www.fnuz.com"}},"_tags":["story","author_picololo","story_227461"],"author":"picololo","created_at":"2008-06-25T13:34:36Z","created_at_i":1214400876,"num_comments":0,"objectID":"227461","points":1,"story_id":227461,"story_text":"Fnuz : A dedicated search engine to free online movies, animes, cartoons and tv-shows.","title":"Free streaming searchengine","updated_at":"2023-09-06T19:17:34Z","url":"http://www.fnuz.com"}],"hitsPerPage":50,"nbHits":13,"nbPages":1,"page":0,"params":"query=PicoJool&tags=story&hitsPerPage=50&advancedSyntax=true&analyticsTags=backend","processingTimeMS":8,"processingTimingsMS":{"_request":{"roundTrip":21},"fetch":{"query":6,"total":7},"total":8},"query":"PicoJool","serverTimeMS":9}