与「LLMs」相关的搜索结果

Updates since then: * Deepseek v4 is out. There *is* a 2-bit quant that can run within 90 GB ( ), and it works, however it's only fast on Apple hardware (I've head ~35 tok/s). On AMD, it's ~7 tok/s. IMO actually taking the effort to properly support more than one hardware manufacturer is a great example of the difference between mere "decentralized AI" and genuine "CROPS AI". I hope we can become better at this. * also has alpha telegram support now. However, the path to adding your account is quite janky * looks promising as a way to run "dense" models (eg. Qwen 27B) more efficiently. It's janky, but on my 5090 laptop it seems to be ~2x more tok/s than llama.cpp * VoxTerm (local AI recording, no third-party servers) continues to be developed And there's a lot more projects coming on the horizon. One other thing that has been on my mind is that there's actually a lot of intersection between "CROPS ethereum access layer" and "CROPS AI". For example, we want a ZK way to make (paid) calls to remote LLMs. But if we have this, then it's just as useful for solving another problem: private RPC reads in Ethereum. Another example: application-specific finetuned LLMs. Leanstral ( ; I get ~38 tok/s on AMD) fits into < 70 GB, but can hold its own against 1T models on writing Lean code. Things like this are a huge boon for writing more secure code ( ). We should have models finetuned for Ethereum-related use cases as well.

Token usage reduction means cutting down the number of text units (tokens) that large language models process, which directly lowers costs, speeds up responses, and makes scaling easier —  @meisshaily #ArtificialIntelligence# #TechNEws# #Tech# #Technology# #LLMs#

🤖 What happens when you hand the same legal case to five different AIs? They disagree. at almost exactly the same rate human jurors do. That's one the findings in the new chapter @federicoast, @williamhwgeorge, and @robertgdean just published in AI and Arbitration (Wolters Kluwer, 2026), "When Decentralised Justice Meets Artificial Intelligence." 63 real Kleros disputes, judged by five frontier LLMs from: ChatGPT, Claude, Gemini, DeepSeek, Mistral. The takeaway isn't which model judges best. It's that you shouldn't trust a monolith AI. Round 2 of the experiment is already in flight: the team is re-running the test on real-world consumer cases from Argentina's Junín pilot and Lemon, where early evidence suggests the AIs and human jurors come to different conclusions on the same cases. Book details below ↓

the best 13 ppl to follow in AI: @karpathy = teaches LLMs @levelsio = AI setup w/ huge ROI @thdxr = honest AI takes @steipete = OpenClaw creator @TakoTreba = marketing queen @aliszu = best AI designs @nico_jeannen = AI ads king @DmytroKrasun = SaaS genius @forgebitz = AI SEO @_lhermann = successful @mattpocockuk = best agent skills @ericzakariasson = composer king @thsottiaux = rate limit reset king if you like this, follow @d4m1n too 🤠

The future of AI and LLMs is not bigger datasets. It is knowing the smallest dataset that forces the whole. My new arXiv paper proves a canonical test for exactly that: What finite fragments are enough to certify the hidden structure underneath them.

Mole 1.5.0 took 6 days, 171 commits, 215 files touched, and 57k+ lines changed. Detailed changelog: 1. Menu bar: added live CPU, memory, and network stats, menu-bar-only mode, right-click quick actions, hotkey support, ghost-state protection, and the new runner animation system. 2. Status: added fan controls for supported Macs, including Auto / Cool / Quiet modes, live RPM, stricter hardware probing, safer restore behavior, and upgrade recovery for old fan presets. 3. Software updates: added update checks and install flows for Homebrew Cask, Homebrew Formula, Mac App Store, Sparkle, and Electron-style appcasts, with clearer progress and safer cancellation. 4. Startup manager: added Login Items, Launch Agents, Launch Daemons, and background item review in one place, with safer authorization behavior so viewing startup items does not ask for admin access. 5. Uninstall: improved alias search, bundle ID search, app metadata matching, Homebrew cask detection, input method discovery, WeChat Input Method support, Doubao Input Method support, and safer root-owned app removal. 6. Clean: tightened log cleanup, protected VPN and proxy app state, guarded Application Support cache cleanup, improved browser and Electron cache detection, and added stronger Trash validation. 7. Analyze: improved disk labels, breadcrumbs, drill-down behavior, folder prefetching, large-directory readability, and trash safety checks. 8. License: improved device management, activation reuse, device reclaim flows, and clearer handling when a license is already used on two Macs. 9. Reliability: fixed Homebrew child-process cancellation, sudo helper reuse, fan preset recovery, MAS inventory edge cases, menu bar ghost states, Startup permission prompts, and release-signing/appcast edge cases. 10. Website and docs: updated the 1.5.0 homepage, release notes, docs, help pages, llms.txt, appcast, and downloadable DMG.

Open-source framework for building real-time voice AI agents! Pipecat is a Python framework for orchestrating audio, video, AI services, transports, and conversation pipelines. Voice-first architecture with pluggable components. What you can build: voice assistants, AI companions, multimodal interfaces, interactive storytelling, business agents (customer support, intake), and complex dialog systems. The framework handles speech recognition, text-to-speech, conversation logic, and real-time interaction. WebRTC and WebSocket transport built in. Ultra-low latency for natural conversations. Why Pipecat: • Voice-first: Integrates STT, TTS, and conversation handling in one framework • Pluggable: Supports multiple AI service providers for each capability • Composable pipelines: Build complex behavior from modular components • Real-time: Low-latency interaction with streaming audio/video Supported services: • Speech-to-Text: Deepgram, AssemblyAI, OpenAI Whisper, Groq, Azure, AWS, Google, and more • LLMs: OpenAI, Anthropic, Gemini, Groq, Mistral, Ollama, AWS, Azure, and more • Text-to-Speech: OpenAI, ElevenLabs, Deepgram, Cartesia, Azure, AWS, Google, and more • Speech-to-Speech: OpenAI Realtime, Gemini Multimodal Live, AWS Nova Sonic, Ultravox, Grok Voice Agent 10.3k+ stars on GitHub. I've shared link to the repo in the comments!

Andrej Karpathy posted the following statement on X: "Personal update: I've joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D. I remain deeply passionate about education and plan to resume my work on it in time."

Personal update: I've joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D. I remain deeply passionate about education and plan to resume my work on it in time.

We post-train LLMs for math, for code, for instruction-following. Why not for scientific discovery? 🫎 MOOSE-Star (ICML 2026) : the first scalable SFT recipe for discipline-agnostic scientific hypothesis discovery. By @Yang_zy223 & @LidongBing from MiroMind.