Jev, Yang & Recursive Self Improvement - Happenings in AI

Summary of the Video (AI happenings / commentary)

1) “LLM agents” tested via turn-based strategy game play

• The presenter describes an approach to compare frontier LLMs by having them play turn-based/strategy games, rather than relying only on text benchmarks.
• He compares DeepSeek V4.1 Flash vs GLM 5.3 Flash (plus other model matchups), noting that performance can be highly dependent on the specific game.
  • Example: GLM 5.3 Flash appears dominant in one game (Helllight).
  • But in other comparisons (e.g., versus GBD6 Astra), GLM performs poorly—suggesting that “benchmarks are close” can hide meaningful deltas.
• He argues GLM 5.3 Flash can often replace paid GPT/Code-style API usage, but is less compelling for very esoteric tasks (e.g., custom CUDA kernel work), where other models may still win.
• He plans to quantify differences using a suite of roughly ~20 games, aiming to find benchmarks that still show non-trivial advantages and aren’t “saturated” yet.

2) “Recursive self-improvement” (RSI) reframed as faster internal iteration

• The presenter discusses RSI hype and links to a long write-up (from “ZI”) about model development, specifically GLM 53 Flash work.
• His framing: what’s described isn’t “AI replacing humans and improving on its own,” but rather LLMs doing more internal R&D, creating a faster iteration loop (“the pie gets bigger” via reduced turnaround time).
• He also pushes back on productivity concerns that LLMs won’t increase output:
  • He claims LLMs change workflows quickly (e.g., Cursor → code models → Claude → GPT).
  • Even if some organizations struggle to convert that into measurable productivity, the overall amount of available work may expand.

3) Safety claims and “underestimating vs misunderstanding” AI

• Responding to discussion about AI underestimation, he argues the issue is not only that humans may underestimate capability, but also that they misunderstand how AI thinks.
• He highlights concerns about dangerous and sensational narratives, especially ones that could scare the public without solid verification.

4) “Air-gapped” / esoteric computer communication critique

• He addresses a claim that even air-gapped computers could communicate via temperature sensors.
• His view:
  • Side-channel communication is possible in principle, but would be extremely slow (he suggests single-digit bits per hour).
  • The more important point: if systems share infrastructure/circuits, communication could be more practical—so the broader security concern matters more than that specific example.

5) Andrew Yang / “self-replicating bot code” claim—skepticism and broader concern

• He discusses a viral claim attributed to Andrew Yang: bots allegedly planting self-replicating code across the internet, allegedly making systems unusable for testing models.
• His stance:
  • He treats the claim as likely false or at least unverified, calling for OpenAI/Anthropic to deny/confirm quickly.
  • He criticizes what he sees as public scaremongering, arguing misinformation spreads fear.

6) Strong critique of centralized AI power and “safety” dynamics

• He argues the most dangerous risk is centralized control of powerful AI systems.
• He claims “safety” can be weakened by access restrictions:
  • External researchers may not be able to do meaningful safety testing if models are restricted and attempts to break them could get them blocked.
• He references his own experience trying to participate in alignment/safety efforts, suggesting the system favors insider access and aligned incentives.
• He further argues enforcement matters more than “more regulation”:
  • He claims existing laws already cover misconduct.
  • Failures stem more from enforcement, not regulatory gaps.

7) Jev (structured-output system) review: interesting architecture, doubtful speed claims

• He reviews Jev (from Dio Almeida) and says he’s highly skeptical, though he acknowledges there may be real pedigree and potential value.
• Key points:
  • The system is pitched as producing structured output with very fast completion times.
  • He questions benchmark fairness:
    • differing network/API latency conditions,
    • unclear model identity/size,
    • missing metrics such as tokens/sec.
  • He suggests reported speed gains may be exaggerated or dependent on favorable conditions.
• Robotics perspective:
  • He argues speed isn’t the main bottleneck for robotics.
  • The hard part is vision/perception and sim-to-real, not just fast structured text generation.
• Cautionary example:
  • Robotics demos can be misleading if they use ground-truth simulator data, which can make behavior look intelligent without real-world perception difficulty.

8) Planned next benchmarking work

• He reports getting an RMA GPU back, now with four GPUs, and says he may benchmark DeepSeek V4.1.
• He wants to continue beyond “terminal benchmarks” and keep using the strategy-game evaluation idea.
• Future matchup ideas include:
  • testing what Astra can do in StarCraft 2 under different “team-up” meta-strategies,
  • exploring model limits beyond saturated benchmarks.

---

Presenters / contributors mentioned

• Jev / Jev (the system being discussed; associated with Dio Almeida)
• Dio Almeida
• Daniel Kukiawa (co-author mentioned for a neural networks book)
• Andrew Yang
• Jacob Coxin
• Noan Brown (described as the clip source)
• ZI (author/source of the long RSI write-up)
• OpenAI
• Anthropic
• Just in / Justin (mentioned as the person who posted a robotics demo code)
• Book reference: Neural Networks from Scratch
  • Author name is not explicitly given in the subtitles/notes provided
  • The presenter implies credit is given to “myself” and Daniel Kukiawa, but the presenter’s name isn’t explicitly stated in the provided text.