Podcasts 2
Latent Space's 92-minute Richard Socher/Recursive interview explores “Eureka Machine” and recursive self-improvement: Socher reports short-horizon agent gains on code and GPU kernels, but says objective design and reward hacking remain unsolved
(episode and transcript)
Explainer →
At the 00:00 opening, the host frames the pursuit of self-improving research systems; the discussion moves through short-cycle benchmark results, guardrails, physical/economic limits and the hard problem of choosing a trustworthy objective. Treat Recursive's performance stories and funding figures as participant-reported, not independent benchmark results.
Training Data's 55-minute Aaron Levie interview describes Box's product-specific agent harness—permissions, search and workflows around models—and why enterprise adoption is slower than coding; his prediction that 90% of enterprise tokens may be agent-initiated within five years is a forecast
(Sequoia episode)
Explainer →
Levie's through-line is that a generic model alone does not understand an organisation's files or access rules: the application layer makes the agent useful. He also questions whether today's token subsidies are sustainable; the diffusion and usage numbers are his view rather than measured sector-wide outcomes.
HF Trending Papers 6
Atria Dawn reports strong science/engineering research-assistant results across 16 benchmarks and 769 participant task records; roughly one-third of assisted tasks were rated infeasible without AI, but the assessment and benchmarks are author-reported (389 HF upvotes)
(paper)
Explainer →
RSIAgent uses an actor, curriculum, verifier and reusable memory to improve agent task performance without changing model weights; reported OSWorld and exam gains should not be confused with an independently verified recursive-self-improving model (65 HF upvotes)
(paper)
Explainer →
Orthrus probes lossless speculative decoding at BF16 precision: exact trajectory matching was only 45% in one comparison, illustrating why faster decoding claims need both numerical and downstream-quality checks (26 HF upvotes)
(paper)
Explainer →
HazardAuditor normalises computer-use agent events before safety review and reports up to a 16.5-point guardrail gain over tested alternatives; the benchmark is a preprint and deployment transfer remains unproven (14 HF upvotes)
(paper) `[RELEVANT]`
Explainer →
When Agents Slow Down studies very long agent runs and finds diminishing gains; splitting budget across multiple attempts beat one long trajectory in its reported Elo comparisons, a reminder that more tokens are not always better (12 HF upvotes)
(paper) `[RELEVANT]`
Explainer →
ScienceBuddy separates a fixed-model inner research harness from an outer reinforcement-learning update, testing scientific problem-solving across four domains; “self-improvement” here has a defined training loop, not autonomous open-ended escalation (15 HF upvotes)
(paper)
Explainer →