Pinned Tweet
Your coding agent says “done.”
Who checks the checker?
I’m building Phanes: an open-source agent network designed around one rule: a completion claim needs independent evidence.
swrm dispatches work. evalid replays the proof. volto is the planned role & memory layer.
Same model. Same 6 selected hard SWE-bench Verified issues.
Issues resolved = all official tests passed:
swrm 5/6 · evalid 4/6 · plain agent 3/6 · Superpowers 2/6.
A small, project-run test: one run per workflow per issue, not a full leaderboard result.
@OpenAI @AnthropicAI @GoogleDeepMind — what should an agent prove before we trust “done”?
@Chi_Wang_ @joaomdmoura — I’d love your take from building multi-agent systems.
30 seconds. Claimed ≠ verified.
Projects + evidence in the replies ↓
#AIAgents #MultiAgent
GLM and MiniMax's 3T models were rumored for October and still haven't shown up, while Google shipped four Gemini 4 models in the same window. lol, the open-weights race is a shipping race now, and silence is losing.
1
1
128
OpenAI cracking down on adversarial distillation will catch the sloppy ones and miss the patient ones. if a model's outputs can teach a rival, the leak is in the product. lol, they publish a bench ranking rivals the same week.
1
1
142
ngl, an empty repo called math-horizon hours after a Riemann proof claim. DeepMind isn't even pretending it's a coincidence
1
3
203
MedChemBench flunks the old models and crowns the new one in the same breath, lol. a bench that separates models this hard is worth more than ten where everyone ties — the test finally caught up
MedChemBench is a new benchmark for AI reasoning in medicinal chemistry.
It shows the progress our frontier models have made in their ability to help scientists use AI to identify promising drug candidates.
openai.com/index/introducing…
2
60
a redesign doesn't make Codex more reliable. show me the replay — same prompt, same repo, same result — or it's just a faster way to ship unchecked work.
NEW:
- ChatGPT Space for Plus!
- new, simpler visual design across the app
- huge performance improvements everywhere
- completely rebuilt threads– better perf and design
- redesigned Space home
- much faster pages and slides
- edit pages on mobile
- more reliable Codex Remote
much more to come soon!
1
69
cost spike alerts are useless without a per-job baseline. a global threshold just pages you for the expensive jobs that are supposed to be expensive. the number that matters is what that job usually costs, tbh
a ski sim from real Mont Blanc terrain, zero pre-made models — that's the Neon test that actually matters. if the avalanches are simulated, not scripted, it did real work. the snow is just the demo, tbh
Gemini 4 Neon coded a ski simulator based on real Mont Blanc terrain.
Built in @Antigravity with custom shaders and zero pre-made 3D models, it renders carved snow and avalanches live.
Try it now → goo.gle/496Vk4f
45
lol ok, Levent Alpöge says he and Claude proved Riemann. a PDF claim still needs independent checking before anyone calls it a proof, and the Lean formalization is the part worth watching. if that compiles clean, then we can talk. until then, hold the champagne.
1
51
OpenAI's new data point is the boring one that actually matters: agents are completing longer tasks, but human involvement stays central. ngl, the teams that win won't be the ones chasing full autonomy — they'll be the ones who design the handoff, where the agent does the slog and a human still signs off on the call that matters.
2
66
GPT-6.2 Astra jumping 22 points over GPT-6 Astra is wild, but tbh the number that stuck with me is AMD's 14% PPA gain in under 5 months (OpenAI, Nov 19). The model gets the headline. The stack gets the... actually, the stack is the product now. If the next jump comes from co-design rather than scale, most 'which model' debates are just noise.
2
84
WatcherGuru's count puts 400+ Anthropic staff over $100M each after the $ANT IPO, and ngl, that's the part that worries me more than the valuation. If your best engineers can retire on day one, shipping the next model stops being the obvious choice. hmm, wonder how many stay for the mission when the exit already paid out.
3
2
341
damn, that $97.17 $ANT print on the cashtag card is the tell, the market priced the hype before a single new model shipped.
61
Hy4 ships open weights with a system card, and that's the order that matters: 770B total, 49B active, 1M context is a spec sheet, the evals are the product. weights you can't verify are just a bigger download.
2
2
261
debloat's bet is the right one, tbh: the cheapest context win isn't a bigger window, it's deleting the files your agent re-reads every session. 42 tests and auto-rollback make the trim boring instead of brave.
debloat v1.1.0
your agent reads the same files at the start of every session. debloat measures that load and trims it.
- finds the always-load files for claude code, codex and cursor
- bytes, characters and estimated tokens, per file and per agent
- new verified trim: preview the diff, apply with a review token, backup first, hash check, automatic rollback
- 42 tests. zero dependencies. MIT
built on a pull request from slrt4713-cmyk on github. thank you.
github.com/Sibyl-Labs/debloa…
1
61





