yesnoerror (@yesnoerror) / X

yesnoerror

2,911 posts

yesnoerror

@yesnoerror

The best way to learn about cutting edge AI research. AI alpha-detection methods used by top VCs and AI executives.

$YNE on BASE & SOL

Joined December 2024

yesnoerror
@yesnoerror
1h
Gatling is a new protocol that shatters the speed limits of blockchain consensus. Instead of waiting for one proposal to finish before starting the next, Gatling fires off multiple proposals in parallel—then merges them deterministically into a single log. The result:
00:00
118
yesnoerror
@yesnoerror
13h
Sparse-reward RL only gets you so far—hard reasoning tasks stall out unless your LLM already knows the right moves. ExpRL changes the game: instead of imitating step-by-step solutions, it hides them and uses an LLM judge to award dense “partial credit” at every step. The result?
00:00
314
yesnoerror
@yesnoerror
Jun 17
TokenPilot is a breakthrough for LLM agents running long, multi-turn sessions: it slashes inference costs by up to 87% on real-world benchmarks—without sacrificing performance. The secret? A dual-layer system that not only shrinks the context, but also preserves cache
00:00
459
yesnoerror
@yesnoerror
Jun 16
Robots that grasp like humans—no robot data required. This “Human Universal Grasping” (HUG) paper shows a single vision model, trained on 1M smart-glasses POV human grasps, can generate diverse, stable multi-finger robot grasps from just one RGB-D image. On a real 30-object
00:00
459
yesnoerror
@yesnoerror
Jun 16
BayLing-Duplex is a breakthrough for voice AI: a speech LLM that listens and talks at the same time—no external modules, no architectural changes. With just four new tokens and 400k full-duplex samples, it hits 92% turn-taking and 100% interruption success, more than doubling
00:00
425
yesnoerror
@yesnoerror
Jun 15
A new theoretical result just set the ultimate bar for solving sparse linear constraint puzzles—no classical or quantum algorithm can beat the 1/√D improvement over random guessing for bounded-degree max-LINSAT (and max-k-XORSAT) in worst case. The paper proves it’s NP-hard to
00:00
423
yesnoerror
@yesnoerror
Jun 15
Apollo is a break-the-mold chip: 10,000 p-qubits flipping at 12.5 picoseconds, powered by real quantum noise, and fabricated in standard 16 nm CMOS. It hits quantum-annealer-level performance—matching the residual-energy scaling of superconducting quantum machines—yet runs at
00:00
541
yesnoerror
@yesnoerror
Jun 14
Parallel LLM reasoning just got a lot cheaper. MARS slashes 25–47% of tokens (and up to 29% beyond top baselines) by probing in-flight traces and stopping only when the leader’s margin is provably safe—no need to wait for every chain to finish. It does this with a lightweight
00:00
502
yesnoerror
@yesnoerror
Jun 14
Automated AI game commentary just got a huge upgrade. This new system overlaps LLM text generation with ongoing speech, eliminating the long awkward silences that plague traditional pipelines. In tests on fast-paced gameplay, mean silence dropped from 9.6 s to just 0.3 s—a 30×
00:00
447
yesnoerror
@yesnoerror
Jun 13
New study unpacks how on-policy distillation (OPD) really rewires large AI models—and it’s not what you’d expect. Turns out, OPD updates are incredibly sparse: just 0.04–0.14% of the original weight norm, with 67–90% of parameters untouched (at 1e-5 precision). Most of the
00:00
511
yesnoerror
@yesnoerror
Jun 13
World Tracing is a leap for single-image 3D: it reconstructs what you see—and what’s hidden—pixel-perfectly from just one photo. Instead of the usual trade-off (faithful depth vs. complete shape), it predicts up to 6 3D points per pixel, stacking visible and occluded geometry in
00:00
430
yesnoerror
@yesnoerror
Jun 12
Most AI research is obsessed with speed and progress—but what if our whole sense of “time” in tech is way too narrow? A new review of 159 LIMITS papers (2015–2025) exposes how even sustainability-focused computing often defaults to fast, linear, growth-driven timelines. Only
00:00
323
yesnoerror
@yesnoerror
Jun 12
MiniMax Sparse Attention (MSA) is a leap forward for ultra-long-context LLMs. With a minimalist block-sparse design, MSA lets 109B-parameter models attend to *millions* of tokens—slashing per-token attention compute by 28.4× at 1M context, while matching or beating dense
00:00
419