@_galyo “Hallucination” is still the wrong abstraction. Frontier LLMs don’t fail because they occasionally detach from truth. They fail because they never had direct truth access to begin with. Transformers are proposition generators, not assertion engines. They interpolate over corpus geometry: • coherence • co-occurrence • discourse priors • token-density topology not external reality. So when a model says something false with high confidence, that’s not necessarily a malfunction. It’s often the architecture operating exactly as designed: maximizing corpus consistency, not world verification. The key distinction: • Assertions require exogenous grounding (sensors, databases, experiments, humans) • Propositions only require endogenous plausibility LLMs only do the second one. This is why “metacognition” alone won’t solve hallucinations. Mapping probability diffuseness → hedging language (“I may be wrong…”) is useful UX, but it’s still an internal statistical reflex inside the same closed system. The map is still verifying the map. Scaling, RLHF, and self-reflection improve discourse discipline, but they don’t create epistemic grounding. The real architectural shift is separation of concerns: Generation ≠ Verification 1) LLMs generate candidate propositions. 2) External systems verify against reality. That’s the missing layer. The future probably looks less like “models that know truth” and more like: • stochastic generators • deterministic verifiers • provenance-aware reasoning stacks • explicit assertion/proposition labeling Not bigger autocomplete. Chaining LLMs doesn’t result in introspection or self-correction. They’re expanded interpolation paths. Longer reasoning traces ≠ epistemology. See the semiotic triad to see the gaps in your current mental model.

SpaceX has just received FCC approval to acquire ~65 MHz of nationwide spectrum from EchoStar for the company's next-gen direct-to-device @Starlink Mobile service. The FCC says the deal gives SpaceX “exclusive-use, contiguous spectrum nationwide” for direct-to-phone connectivity from orbit. Next-gen Starlink Mobile is going to be incredible, enabling 5G speeds from space in the middle of nowhere.

From a pure "do good for the world" mission perspective, having the acting like a solid personalized tutor is one of the better uses of AI. If OpenAI cares about the mission of making the world a better place, tutoring should be an area of investment, not one to silently remove. https://t.co/hC1ltmETJ8

+2 more

Grok Voice Think Fast 1.0 ranks #1 on the Artificial Analysis τ-Voice benchmark for real-world agentic customer service resolution Absolutely outperforming GPT-Realtime-2 (High) and Gemini 3.1 Flash by a huge margin That's a massive 12%+ lead over OpenAI's best model that just released a few days ago Grok is running real-time background reasoning without the latency penalty, which is why it is already handling live Starlink phone operations autonomously at scale

A few community stories we loved 👇 🧵 Explore the full community feed at https://t.co/fhKDzlntJj and tag us in your creations.

Peanut's Floating Problem: what happens when your pet elephant can't stop hiccuping, and starts to float with each hiccup? School picture day is tomorrow, and the hiccups show no sign of slowing down: https://t.co/hmOob3N21o

The Last Resume: you are the final human career advisor in a world governed by cold, predictive algorithms. Suddenly, a glitch ripples through the global workforce system, flashing a single, impossible job opening on your screen: 'Human Role: Undefined': https://t.co/Yrezx2XTKI Shoutout to @ilovegarick!

Bolt's big dark: your best friend Bolt is a small silver robot with one wobbly antenna and a tiny light on his chest that blinks when he's nervous. He says the dark feels too big and too quiet and he doesn't know what's in it. Bedtime is in 10 minutes, and it's up to you to reassure him that everything will be okay: https://t.co/Sc4OOwLMvJ

Craft your own at https://t.co/rNUKOjyRdx and tag us when you share - we'll send you swag!

@AnthropicAI @emergentlabs Who's next? https://t.co/Kpzpj2UwQR

@AnthropicAI @emergentlabs Learn more about Emergent + Claude Platform on AWS: https://t.co/n3VajMRVCu

InstaAgent (@InstaAgentAI) helps B2C companies scale social media marketing across hundreds of personas. They’ve already reached $1M ARR in just 10 months. Congrats on the launch, @klwongkyle & @tseungcolin! https://t.co/wnI34gS0Oo https://t.co/CFmHkvyS81

Not every meeting should be an email. Sometimes it should be spatial. Join the Public TestFlight and let me know what you think! https://t.co/Gs0A1KC366 #visionOS #SpatialPersonas #TestFlight https://t.co/dEhT0xIAJe

Introducing Googlebook, the first laptop designed for Gemini Intelligence. It’s crafted for heavyweight performance, built with Gemini at the core and perfectly synced with your Android phone. Coming this fall. 💻✨ #TheAndroidShow https://t.co/rn4pztApmp

Today, we introduced Gemini Intelligence, which brings the best of Gemini to our most advanced devices. Gemini Intelligence integrates premium hardware and innovative software to help you stay a step ahead and work proactively to get things done throughout your day. #TheAndroidShow

Introducing PROWL! We’ve built RL agents that explore game environments, tasked with discovering failures in world models across physics, visuals, and actions. Those failures then become training data in an automated loop that advances world model performance. https://t.co/LeCzIxfl2Y

just launched: @mobbin mcp 💛 your AI agents can now search 600,000+ real app screens. paywalls, onboarding, checkout, permissions — from apps that already shipped. ai tools can write code. they just don't know what good looks like. now they do. 👉 https://t.co/WL3IofRU84 https://t.co/butDiFPJbA

parameter golf was a blast. 2,000+ submissions. 1,000+ verified github accounts. ideas ranging from quantization and depth recurrence to TTT LoRA, SSMs, H-nets, JEPA, and more. autoresearch made iteration dramatically faster — and led to emergent bulletin boards, issue threads, unofficial leaderboards, and agent-built writeups that helped everyone learn from everyone else. it felt like a glimpse of where interaction with AI is headed: humans setting taste and direction, agents helping explore, coordinate, and share what works. our goal was simple: make ml research accessible to anyone, anywhere. it was amazing to see that happen. full recap: https://t.co/FxvcbImWzL future events: https://t.co/qanojnkjmJ

PyTorch 2.12 introduces updates across compilation, distributed systems, export, graph capture, and accelerator support. Join Andrey Talman (@Meta), @albanDesmaison (@Meta), and @joespeez (@reflection_ai), moderated by @Chris_AI_HPC (@Meta), on May 20 at 10:00 AM PT for a live Q&A covering the release and answering questions from the community. Highlights include the new device-agnostic torch.accelerator.Graph API, up to 100x faster batched eigenvalue decomposition on CUDA, support for microscaling quantization formats in https://t.co/DQ6E1nOVvb, fused Adagrad optimizer support, FlightRecorder updates, multi-GPU and multi-node profiling improvements, updated backend selection for torch.linalg.eigh on CUDA, and expanded CUDA, ROCm, XPU, MPS, and Arm platform support. 🔗 Register today: https://t.co/LThdjj96F0 #PyTorch #OpenSourceAI

We're excited to release Medmarks v1.0 + a technical report! This is an update to our Medmarks benchmark suite, the largest open-source automated suite for evaluating the medical capabilities of LLMs. We added 10 benchmarks (20→30) and 15 models (46→61) to the leaderboard! https://t.co/QAhQtzhRF5

Today's Hermes Agent Jam starts in 90 minutes! Join the Nous team in our Discord to share what you're working on @ https://t.co/avq295YPIm https://t.co/PLMoV6fwLK

Improvements to Perplexity Computer's traceability: 1. In finance tasks, hover over any number to see where it was pulled from (disclosed vs. derived) 2. Realtime market data (e.g. market cap) includes a timestamp of when it was pulled 3. Full calculation trace for derived values shown in-thread and in sidebar 4. Tap to audit filings for each value, pre-scrolled to the correct page 5. Or tap to visit Perplexity Finance for a deep dive on other company information

For a driver born without arms, FSD Supervised is life-changing accessibility “I was born without arms and have driven with my feet my entire life. I’m a fully licensed driver, and traditionally I drove with my left foot on the steering wheel and my right foot handling the gas and brake. My only legal restrictions are automatic transmission and power steering. Over the years, though, the strain from my congenital birth defects has led to significant arthritis in my hips. I drove a Model 3 for the past seven years, and it honestly helped extend my independence in a huge way. Recently upgrading to the Model Y – along with Full Self-Driving – has been a complete game changer for me. It dramatically reduces the physical pressure and fatigue of driving and has helped preserve a level of freedom and mobility that means a great deal to me. Most people understandably think of Tesla in terms of innovation or sustainability, but for some of us, this technology truly becomes life-changing accessibility.” – John F.

Today we launch Agency. Instead of prompting AI, AI prompts you. It watches your tools, finds useful work, drafts the action, and asks for approval. One click, and it sends the reply, creates the video, does the fix, solves your distribution, monitors your alerts, orders food, cancels subscriptions, finds the best product and sends screenshots for confirmation. It runs 24/7 from Telegram with thousands of one-click integrations like gmail, slack, DataDog, Linear... You can use it with your Claude Code / Codex subscription. And because the app and agent are on the same machine, Agency can edit and improve itself based on your decisions. Reply "Agency" and I will send you the link.

The 2026 PyTorch Docathon is underway, bringing the community together to improve the documentation developers rely on across PyTorch, Tutorials, and ExecuTorch. Contributors have already closed 100 pull requests and 73 issues. The Docathon runs through May 17, so there's still time to contribute. 🔗 Learn more and register: https://t.co/dvqqy28fEM 🔗 Track the leaderboard: https://t.co/bMxzRvrEfV #PyTorch #Docathon #OpenSourceAI

Today we're releasing Marionette: create robot motion with your hands (Reachy Mini). Browser-based, works on phone or computer. Hold the robot to record a move, add sound, share it with the community. Same format as our emotions dataset! An open source robot whose movements are community created. Let's go!

Step 3.5 Flash from @StepFun_ai is now free again in Nous Portal for the next 15 days! https://t.co/XcH8pUxrqK

Step 3.5 Flash from @StepFun_ai is now free again in Nous Portal for the next 15 days! https://t.co/XcH8pUxrqK

respect to whoever decides train/global_step deserves a panel. someone has to track the steps. fit bit global_step. credit: @frances23398579 https://t.co/RpmuHEOnz3

time to see who's diverging https://t.co/L8X6MBEiMc

How many tokens can we rip on this commute? https://t.co/q1JjOaTpB1

AI is becoming a research partner for scientists around the world. But as more researchers rely on AI for experiment design, reviews and analysis, an important question emerges: what happens when trust shifts from colleagues to machines? Science advances through verification, debate and human judgment. Those foundations still matter in the AI era. https://t.co/dMa1IRr6RP @ConversationUK @ConversationUS