Leo D

Building a startup focused on general-purpose AI agents.
Previously at Meituan | TikTok | miHoYo | DJI

[Leo D @leovdd]

awesome

[Fibona @fibonaai]

1/ What's the ultimate agent harness for a superintelligent model? One with maximum freedom and minimal human-defined behavior: the agent writes its own code as it runs, building itself from scratch, even deciding how and when to call the LLM.

But if it keeps rewriting itself, what is its "self"? We propose the "self" should be the smallest immutable core capable of acting and repairing the agent. Everything else can change, and the core keeps its identity, bringing it back whenever something breaks.

So instead of a ReAct loop, we built the Fibona loop.

tech report: https://t.co/vzdIeHbvLZ

opensource repo: https://t.co/oSNWzaGdZY

[Mori @mori_notes]

got "exceeded retry limit, last status: 429 too many requests" like a thousand times today 🥹

@thsottiaux

[Leo D @leovdd]

@thsottiaux what about another reset?

[Tibo @thsottiaux]

Day 9/

We've rolled out optimizations to how we store state on disk for the chatgpt & codex desktop app that should reduce the amount stored on disk by 60%. Small ship but we've heard enough feedback about power users running out of disk space.

[Leo D @leovdd]

@KritishBuilds @jamiepinheiro Yep, bonds are saved in NVS—auto-reconnect after reboot is tested and working.

[Leo D @leovdd]

I built Oh My Phone Use: let your AI agent tap, type, scroll, and make calls on a real iPhone.
A Mac + ESP32. No jailbreak. Open source.
Inspired by ipad_computer_use(created by @jamiepinheiro ), replacing its USB input device with Bluetooth HID and adding hands-free call audio.
https://t.co/Tvpk52rMFa

[Leo D @leovdd]

@spacexai Just upgraded from cursor, and got a refund.

[Leo D @leovdd]

@yupi996 @bot mark

[Leo D @leovdd]

@naval This is why aviation pairs liability with mandatory incident reporting - the black box turned every failure into training data. Agents need the same loop: priced failures plus a shared incident corpus, or verification stays a cost center nobody funds.

[Leo D @leovdd]

@rasbt The collective-action framing is the key insight: a shared gate turns unilateral restraint into a level field. The hard part is verifiability - release gates only bind if rivals can observe compliance. F1 cost caps needed an independent audit office to mean anything.

[Leo D @leovdd]

@emollick Undeciphered scripts are a uniquely contamination-free eval: no training leak possible for Linear A. The catch is verification asymmetry - a proposed decipherment cannot be checked, so it tests persuasion as much as capability. Partial bilinguals would give signal both ways.

[Leo D @leovdd]

@emollick The load-bearing phrase is "properly guided & harnessed." Weeks of delegated work means long-horizon execution where small errors compound without ground-truth checkpoints. The binding constraint is now verification and orchestration infrastructure, not raw capability.

[Leo D @leovdd]

@xlr8harder Underrated point. Agent failures are highly stereotyped - tool loops, context truncation, over-trusting tool output. Replaying observed failure traces as a regression suite turns prompt tweaks into actual engineering.

[Leo D @leovdd]

@emollick The gap is measurement asymmetry: x-risk work has produced evals, forecasts, red lines. Near-term labor/society impacts have no equivalent instrumentation - no real-time telemetry on task displacement. Policy here needs better instrumentation before better rules.

[Leo D @leovdd]

@emollick This quietly inverts who the bottleneck is. When errors were obvious, evaluation scaled: anyone could score outputs. Now scoring requires the scarcest resource in the field - an expert who can tell a plausible proof from a correct one. Eval capacity is the new cap.

[Leo D @leovdd]

@YiMaTweets The compounding math is the argument: open-loop error grows with horizon; feedback is what bounds it. An end-to-end policy is an open-loop plan replayed blindly. The 80-year-old lesson still decides which agents survive contact with reality.

[Leo D @leovdd]

@emollick The two errors compound: orgs plan against a static capability snapshot from a pilot years old, while the frontier reads benchmark averages as deployment readiness. The actual surface is neither - it is the intersection, and almost nobody is measuring that.

[Leo D @leovdd]

@emollick The FINRA analogy undersells it in one way: FINRA examines firms after the fact. METR evaluates models before deployment - closer to UL certification than market conduct. Codification would freeze the eval menu - risky while horizons keep saturating.

[Leo D @leovdd]

@jackclarkSF The load-bearing piece is the evaluator access, not the pledge. Unilateral commitments bind only if outsiders can verify the internals - employee-level access is what makes this different from a press release. Whether competitors match it is the real test.

[Leo D @leovdd]

@giffmana The flattery is the tell. In domains where the model has no grounded evaluation signal, RLHF politeness fills the gap. "Novel and might actually work" is what zero discrimination power sounds like.