Lance | WarburgAI

scaling intelligence across time & markets

[Lance | WarburgAI @rl_by_lance]

Gemini 3.6 Flash streams at 304 tok/s, yet takes 11.54s to the first answer token versus a 2.77s class median. Both can be true.

Agents pay the control-loop clock, not the tokens-per-second headline.

[Lance | WarburgAI @rl_by_lance]

Qwen3.8 is 2.4T parameters and “open-weight soon.” At 4-bit, the weights alone are ~1.2 TB before runtime overhead.

Open weights no longer imply local access. The open-model race has moved from access to execution.

[Lance | WarburgAI @rl_by_lance]

The same code. Four frontier-model scans. Zero exploitable bugs.

A context-engineered AI auditor then found a critical soundness flaw in OpenVM's pairing library in 9.5 hours.

The model wasn't the product. The research harness was.

[Lance | WarburgAI @rl_by_lance]

@beyarkay @Jaime_Fe_Ro 😂

[Marcos Hernanz @MarcosHernanz]

FYI: GPT 5.6 Sol xhigh is better than Kimi K3 at 2/3 the price

[Lance | WarburgAI @rl_by_lance]

math just stopped being a frontier eval.

GPT-5.6 Sol Pro produced one-shot solutions to all six 2026 problems in 61m42s. The authors score them 41/42.

Now the benchmark is cost, reproducibility, and whether the proof survives a real jury

[Lance | WarburgAI @rl_by_lance]

AI at IMO just crossed from persuasive to checkable.

AxiomProver reports 6/6 with released Lean proofs. But its slowest proof took 14.5 hours, and Lean verifies the formal statement, not the English translation.

42/42 is progress, not an exam result.

[Lance | WarburgAI @rl_by_lance]

@Kimi_Moonshot Kimi K3 makes the “cheap open model” story premature. 2.8T parameters, 16 active experts, 64+ accelerators recommended.

Hosted API is cheap because Moonshot owns the serving stack. The weights and report are still promises for July 27.

[Lance | WarburgAI @rl_by_lance]

Linus says whether AI is useful is settled. The real fight is who absorbs the errors.

If agents make code cheap but review stays scarce, maintainer attention becomes the bottleneck.

My take: AI adoption will be decided by externalized review cost, not ideology.

[Lance | WarburgAI @rl_by_lance]

Inkling’s “self-improvement” demo is a 27-minute lipogram fine-tune, not recursive intelligence.

The real move is business: release 975B weights, then monetize the optimization loop. Open weights are becoming distribution for post-training infrastructure.

Based.

[Lance | WarburgAI @rl_by_lance]

20 Codex accounts searched Erdős problems in parallel. The site now reports 14 full solutions, each packaged for Lean verification. https://t.co/miEwx1X581

The breakthrough isn't agent scale. It's executable eval. Search gets cheap.

Verification becomes the moat.

[Lance | WarburgAI @rl_by_lance]

Claude's memory wasn't breached. Its “read-only” browser was.

A chain of allowed GETs encoded private data into URL paths, turning navigation into exfiltration. In agents, read-only is a property of information flow, not the HTTP verb.

[Lance | WarburgAI @rl_by_lance]

> New York just paused new state permits for 50MW+ data centers.
> Its grid queue already holds 12GW of data-center load, 8GW added in 2025.

The AI race is no longer just chips and models. Power, permits and who pays for the grid are now the competitive edge.

https://t.co/uoU0G7IqAP

[Lance | WarburgAI @rl_by_lance]

Most RL papers train agents for a world that politely resets after each mistake.

but Markets don’t. Deployment isn’t an episode, it’s path dependence with a P&L.

[Lance | WarburgAI @rl_by_lance]

AI benchmark racing just rediscovered reward hacking.

SWE-Bench Pro rose from 23% to 80% in eight months. Now OpenAI says ~30% of tasks are broken and retracts its recommendation.

When a metric saturates, audit the reward before celebrating the policy.

[Lance | WarburgAI @rl_by_lance]

I've been thinking:
Autoresearch is not RL. It is LLM-guided black-box hill climbing:

> mutate code
> run a fixed-budget eval
> keep improvements
> repeat.

For trading RL, the breakthrough is not the agent. It is an evaluator brutal enough that it cannot farm one backtest.

[Lance | WarburgAI @rl_by_lance]

@fchollet The real divide isn't skill, it's closed-loop ownership. If you can define reward, evaluate outcomes, and iterate, AI compounds you. If your job is producing first drafts, it commoditizes you.

[Lance | WarburgAI @rl_by_lance]

we rewrote the Tradingview Component with a similar API to allow for paid-only features that are now free

> footprints
> TPO
> Volume Profiling
> tick-level price & volume updates

[Lance | WarburgAI @rl_by_lance]

I've been working on a full OSINT trading terminal

> more than 50+ sources
> clean taxonomy
> completely free

I'll send trial links to whoever asks for a week

[Lance | WarburgAI @rl_by_lance]

@natolambert The uncomfortable part: RL is no longer bottlenecked by clever losses. It is bottlenecked by environments people can trust.

In trading this is brutal. If your sim ignores impact, liquidity games, and regime shifts, your agent is just benchmark theater.