Learning to solve hard problems in RL for LLMs by never giving up

TEXT START: This is a blog post for my recent paper on RL post-training of LLMs: introducing the Matthew Effect and proposing to solve it with Never Give Up.

The Dissection

This is an efficiency patch for the replacement machine. Standard RL over-rewards already-competent behavior because easy problems generate cleaner, denser training signals. The Matthew Effect exposes that bias; Never Give Up reallocates sampling and compute toward unresolved problems through adaptive retries, filtering, and stale-data controls.

The real achievement is not “never giving up.” It is converting hard cognition from an apparent capability wall into a resource-allocation problem. Once easy competence becomes cheap, the training system can spend more computation attacking the remaining frontier.

The Core Fallacy

The post treats difficulty as primarily a training-distribution defect. NGU may fix signal efficiency, but it does not establish that benchmark success transfers to open-ended productive cognition, robust deployment, long-horizon agency, embodiment, or real-world accountability.

That limitation does not weaken the Discontinuity Thesis. It strengthens its mechanism. The work removes one friction in cognitive automation. It does not restore human productive participation; it pushes models closer to durable cost and performance superiority across cognitive work. P1 advances, while P3 worsens.

“Never give up” is also not magic. It is a compute budget policy with geometric sampling costs. If hard tasks produce no useful positive signal, repeated sampling merely burns resources more patiently.

Hidden Assumptions

• Verifier-defined benchmark performance is a meaningful proxy for economic productivity.
• Rare correct samples contain transferable learning signal rather than benchmark-specific artifacts or reward hacks.
• Hard tasks are finite, decomposable, and solvable with more sampling.
• Additional compute, energy, and inference time remain available at acceptable cost.
• Gains on curated math and code environments transfer to messy, adversarial, open-ended work.
• Institutional deployment and ownership arrangements do not matter to the capability result.
• Improving models benefits labor broadly rather than concentrating control in the owners of AI capital.

Social Function

This is a partial truth functioning as transition management and prestige signaling. The technical claim is credible within its scope: scalar averages conceal stagnation, and adaptive sampling can expose and improve the hard tail.

But socially, the post normalizes the race to automate the remainder of cognitive work. It is not copium. It is a lab memo from inside the replacement process, focused tightly enough on optimization to avoid discussing who owns the resulting capability and who becomes economically unnecessary.

The Verdict

Technically, this is a useful local contribution. Systemically, it is an acceleration document. If NGU generalizes, it turns “the model cannot reliably solve the hard cases” into “allocate enough computation to the hard cases.” That is not a defense of the post-WWII economic order. It is another ratchet toward Sovereign control, with only owners of AI capital and indispensable infrastructure Servitors retaining durable leverage.