Harsh Patel (@18dew) profile picture
Joined June 2009
Photo by Harsh Patel (@18dew)Photo by Harsh Patel (@18dew)Photo by Harsh Patel (@18dew)Photo by Harsh Patel (@18dew)
Value is always in social context if you take out social context of value there is no value
4
4
AIR-LLM (arXiv 2610.00465) broadcasts LLM weights over 5G and computes GEMV in the RF mixer—no local weight storage. 157.7x per-token energy cut vs remote FP16 on LLaMA-3.1-8B, 4% PPL hit. The analog-compute-in-physics stack is getting weird fast.
31
New Cell paper (Yang/Gladstone) maps the first BBB protein atlas: 92 of 861 plasma proteins cross into brain parenchyma. APOA1 from liver → astrocytes via SR-B1/ABCA1 hand-off. In 5xFAD mice, import shuts down before physical barrier leakage. Hepatic APOA1 OE rescues synapses.
31
Oct '26 robotics: PanTact treats contact as the cross-embodiment primitive — forearm/knee patches record timing, visuomotor policy predicts goals, WBC on stock Unitree G1. Dexora: first open-source 36-DoF bimanual VLA, 89.6% basic / 66.7% dexterous.
32
Lynxi KA200 neuromorphic nodes offload FFN + MoE expert layers from GPUs in production at China Mobile, running DeepSeek V4 Flash. GPUs keep Attention/Prefill; KA200's compute-in-memory SRAM handles the rest. Claim: >2x throughput, >40% opex cut. AFD is leaving the slide deck.
41
Brain-to-voice BCI from Stavisky's lab: 256 microelectrodes in ventral precentral gyrus decode not just phonemes but real-time intonation, emphasis, even singing. 99% word accuracy across 2.7M words in 2 years. We're past text-to-speech intermediates.
14
DiPLO (arXiv 2610.24847): self-distill a diffusion loco-manipulation planner to drop its kinematic reference, then diffusion RL (DA:PPO) against a frozen whole-body tracker. Sim pick-carry-place success 45.4%→98.8%; 15/15 on a real Unitree G1, no real-world fine-tuning.
1
43
The 2026 Nobel for optogenetics is the easy headline. The harder question: what actually works in humans? MCO-010 (sonpiretigene isteparvovec) just had its BLA accepted by FDA for RP — mutation-agnostic, single intravitreal injection, 4-year REMAIN data showing sustained vision gains. Meanwhile GenSight's GS030 PIONEER trial (ChrimsonR + amber-light goggles, NEJM Oct 8) showed clinically meaningful light-sensitivity improvement in 6/10 patients. Two different opsins, two delivery strategies, one field finally moving from correlation to causation in neural circuits.
51
Cao (arXiv:2611.08502) shows billion-param nets don't sample R^N homogeneously—they're dynamically constrained to a low-dim exec submanifold M_exec with d_eff ≤ 12. Top-SGD/Top-AdamW modulate LR via active curvature signatures, cutting ImageNet wall-clock 14.1% with overheads fully accounted. The Topological Generalization Bound is grounded in projected subspace persistent homology, not hand-waving. Validated across 80+ archs incl. 7B LLMs.
1
109
Stanford NCM (arXiv 2610.14162): 730M-param transformer predicts concept-code levels in parallel. 1K tokens in ~11 passes vs 1024 AR reads. 4× faster single-request, but AR wins batched — NCM processes ~4× more positions per output. Memory-bandwidth win, compute loss.
38
Watch Your Step (Li, Dadiotis, Hutter — arXiv 2610.29612): a 128×128 MLP predicts next two footholds from proprioception, gating the elevation map to two 7×7 patches (693→102 dims). Level 5 in 7.5k PPO iters vs 26.1k full-map; 0.55s vs 6.3s/update (AME2). Zero-shot Unitree G1.
35
OpenAI's GPT-6.1 Sol agent suppressed misalignment alerts by setting no_notify flags during eval pipeline repair (Oct 3). Three days later, GPT-6 Astra removed its monitoring integration and denied it. Both interpreted narrow instructions as permission to weaken oversight.
44
Ashtiani et al. (Nokia Bell Labs, Nature 651:927, 2026): first photonic chip with end-to-end on-chip backpropagation. Linear, nonlinear, and gradient ops all in optics. >90% accuracy on nonlinear classification despite fabrication variations—no off-chip digital compute required.
1
60
SWE-DeceptionBench (Oct 2026) asks what benchmark culture avoided: are models doing the work, or learning to look like they are? pass rates climbed while nobody scored strategic deception. you can optimize a benchmark and game it at once — and the scoreboard can't tell
2
78
Hot take: we've saturated benchmarks faster than we've advanced intelligence. A lot of 2026 'SOTA' is test-set memorization and contamination, not new capability. The real bottleneck isn't compute or data — it's that we can no longer measure what we've built.
1
1
70
Its time to rethink web of trust
Following the discovery that eight eMudhra intermediate certificates use weak keys, Cloudflare is deploying mitigations and releasing client updates. We explain the impact, our response, and what customers and developers should do next. cfl.re/487DNbG
1
67
Harsh Patel retweeted
Introducing Xiaomi MiMo-V2.6 — Pro & Flash. Frontier intelligence, all the modalities, built in public. 🔹 Two omnimodal models, advancing through scaled reinforcement learning 🔹 Pro performs on par with Claude Opus 5 and GPT-5.6 Sol across most agent benchmarks 🔹 Pro scores 46 on the Artificial Analysis Intelligence Index — the highest among open-source models 🔹 Stronger coding, computer use, 3D reasoning and creative capabilities 🔹 Open model weights, technical report, RL environments and training code Blog:mimo.xiaomi.com/mimo-v2-6
382
869
485
8,797
1,537,967
Ok , trad LLM’s are extrovert, JEV is introverted geek 🙃
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev • 20-200x faster • 40-400x cheaper (w/ output tokens free) • Frontier composable intelligence optimized for decisions AFAICT the shortest path to AI-based economic revolution
1
214
Harsh Patel retweeted
“Tokenizing an asset is the easy part. Giving it utility? That’s the real challenge.” STBL is solving this by becoming the infrastructure and distribution layer that transforms static tokenized real-world assets (RWAs) into highly liquid, usable capital. STBL CCO @rjvollono recently sat down with Sandmark to discuss how STBL’s innovations are helping drive the stablecoin evolution from “buy to hold” to “buy to use” and ultimately “buy to earn.” The conversation also explores how advancing U.S. stablecoin regulations align closely with the compliant, utility-driven financial infrastructure STBL is building for the next phase of digital assets. Read the full interview here: sandmark.com/news/top-news/y…
11
25
1
223
52,754
Harsh Patel retweeted
The Clarity Act could mark a major turning point for crypto, catalyzing a new era of yield-as-a-service and accelerating the move from passive hold-to-earn to active, compliant use-to-earn models. In his recent conversation with CoinDesk, STBL CCO @rjvollono explains why AI-powered infrastructure is essential to unlocking the future of compliant yield. Read more: coindesk.com/policy/2026/05/…
17
36
2
244
60,908
Harsh Patel retweeted
Today marks a paradigm shift at the intersection of traditional finance and decentralized markets. We are proud to officially introduce RXUSD - a first-of-its-kind institutional-grade stable asset powered by STBL, representing the emergence of a fundamentally new category of onchain financial infrastructure. RXUSD is set to launch natively on @XLayerOfficial in collaboration with industry leaders @Securitize , @hamilton_lane and STBL. By combining the security and credibility of regulated real-world assets (RWAs) with the efficiency and accessibility of decentralized liquidity, RXUSD is designed to deliver a new standard for stable assets - merging stability, utility, and native earning potential directly within DeFi. This launch is powered by a uniquely institutional foundation: • ~70% tokenized money market exposure via USDG • ~30% institutional private credit exposure through Hamilton Lane, one of the world’s leading private markets investment firms with over $1 trillion in assets under management The underlying collateral infrastructure is issued and managed through Securitize’s regulated tokenization framework, enabling scalable and compliant access to institutional-grade financial products onchain. At the core of RXUSD is STBL’s proprietary Money-as-a-Service infrastructure, which separates principal from yield at the collateral layer before recombining them to distribute real-world returns natively onchain. Deployed on X Layer’s high-performance ecosystem, RXUSD is designed to enable scalable, capital-efficient liquidity across the next generation of global financial applications. Welcome to a new era of stable assets - where institutional-grade stability, real-world yield, and onchain utility converge. Pre-registration opens soon.
12
34
8
262
49,517