djev is def not like jev... haolab larpsniped😭
🚀 We let Jev play Minecraft. And we just can't beat it! 😭
⚡ Jev: 24 ms decisions
🧠 You: ~200 ms reactions
It moves before you even see it move. Way too strong!
🎮 mc.alexzms.com (join the server to win)
6
1,120
hot path is the only justification for this and is such low loc relative to everything else in a program not v plausible/frequent. also models would make the most sense for understanding compiler flags so gains seem small
Rust is a good prompt compilation target for the moment, but so is C++. And soon assembler. Then microcode. Myopic to think we're going to stop the agentic drill bit until it reaches computing bedrock.
1
6
623
ai labs discover Promise();
Google launched Gemini 3.8 Live. NVIDIA dropped new research on NemotronLabs VoiceChat. Meta expanded business calling through Muse.
3 of the biggest voice AI updates this week. Full roundp 👇
• Google: Gemini 3.8 Live can keep responding while a tool finishes its work. Devs can adjust reasoning through the Extended Thinking version.
• NVIDIA: NemotronLabs VoiceChat combines full-duplex speech with tool calling and reports 82.5% tool-selection F1 on Full-Duplex-Bench 3.0.
• Nuance Labs: $50M Series A led by Lightspeed to fund a model that handles audiovisual input and output together. Research preview expected later this year.
• Meta Muse: US business calls are rolling out to more beta users. People who asked for it earlier get priority.
• Instinct: Concierge is reaching early users with tasks like booking reservations by phone, requesting dentist cancellation slots and sorting out cable billing issues.
• Speechmatics: Agent STT uses the Linden 1 model to better catch names, account numbers and short replies.
• Deepgram: India endpoint is now GA. The new Nova-3 Pharma model focuses on drug names and pharma terms.
• SpaceXAI: Grok Voice Transcribe 2.0 does batch transcription at $0.10 per audio hour and streaming at $0.20.
• AssemblyAI: The new Dictation API turns speech into cleaned-up text, with Blurt as an open-source dictation app.
• DeepL: Speakers keep their own voice when translating across 12 languages. Meeting translation is out of beta on desktop. External browser access is still in testing.
• Pindrop: BotStopper can now be bought on its own to flag automated callers. Its AI Voice Consortium registry is past 5,000 voices.
• Retell: Conductor now works across the whole dashboard and can dig into calls using info from your workspace. Users approve any production edits it proposes.
• ServiceNow / Krisp: ServiceNow is using Krisp's VIVA SDK for voice agents handling IT, HR and customer service calls.
• smallest.ai: Bolna (YC F25) users can now pick Pulse for STT and Lightning for TTS.
• LiveKit: Eligible startups get cloud and inference credits plus waived plan fees worth ~$23K, across 7 inference providers.
•Cekura: We launched our latest STT benchmarks, check them out at benchmarks.cekura.ai/stt
The use cases are getting really specific, like recognizing medication names or calling a restaurant to book a table.
At @cekuraAi we test how voice agents handle what goes wrong along the way. Every new capability means new things to test, like a misheard prescription or a booking that needs human approval.
1
581
the robot china argument also applies to USA 1.4b ppl > 800m + robots > 400m + robots
hence we need goood robots
2
400
why not open source noc mesh 👀
MeshKV: A Network-on-Chip KV Cache Fabric for Scalable Transformer Decoding Accelerators
Dong Liu, Yanxuan Yu
arxiv.org/abs/2609.19207 [𝚌𝚜.𝙰𝚁 𝚌𝚜.𝙰𝙸]
ALT Autoregressive transformer decoding is constrained by irregular key-value (KV) cache movement on tiled accelerators. Prior compression and DRAM-placement systems still concentrate traffic on centralized memory paths that bottleneck long-context serving. We present MeshKV, a KV cache fabric that moves blocks as packetized flows over a lightweight NoC. It co-designs (i) TaKV affine striping to spread homes and cut hotspot load, (ii) Mare multicast with verified duplicate suppression, and (iii) Pad, which overlaps prefetch, tile multiply, and streaming softmax behind credit-aligned FIFOs. Together they convert bisection back-pressure into useful KV transfer. On our 8x8 FPGA implementation with LLaMA-2-7B and Mistral-7B at 8K-32K, MeshKV reduces interconnect traffic by up to 58%, improves KV bandwidth utilization by 2.1x, and delivers up to 1.9x multi-stream throughput.
4
894
