Pixel

@carbonyxxx He opens the course by telling them it will never make sense. That’s the only honest way to teach it :D

Ramamurti Shankar, Yale physics professor: "Yale pays me $400K to teach quantum mechanics - and I open every course by telling students it will never make sense. Reality doesn't run on your intuition. Trust the math or stay wrong." most teachers try to make the hard part feel intuitive. Shankar does the opposite: he warns the room it cannot be made to make sense. "there's no way to make it reasonable. it's not a reasonable world out there. I can only tell you what it is." the quantum world doesn't run on everyday logic. "this is not daily life. strange things happen." your intuition isn't just weak here - it's the wrong tool. at that scale you don't get certainty, you get odds. a particle has no definite position waiting to be found - only a wave of probabilities. the outcome isn't there until you look. "it makes sense to me because I've seen it," he tells them. "I have no clue how it sounds to you." he won't promise it'll feel right. only that it's true.

@Di_Krass_ The math nobody respected became the native language of every model that matters))

Gilbert Strang, MIT mathematics professor: "MIT paid me for 60 years to teach the math nobody respected. Matrix was a dirty word. Now every neural network runs on it. It's the plain math nobody wanted to learn, until it turned out to run the future." the subject he spent his life on was once dismissed - mathematicians looked down on linear algebra as beneath them. then it quietly became the engine of the entire AI era. every neural network, every model like ChatGPT, is just matrices multiplying matrices - the exact math Strang made simple enough for anyone to learn. same story i keep telling: the foundation is never the flashy part. it's the plain math nobody wanted to teach, until it turned out to run the future. "this is an unforgettable day for me," he said. the last lecture of the man who taught the machines their native language.

@pulmencr Described the desire. Got the full signal-processing pipeline back. Zero lines of math written by hand.

I fed a blackbird's song into Claude Opus 5 and asked it to build a real-time 3D spectral visualizer This is what came back, and it looks like a spaceship interface Red spheres pulsing across a 3D acoustic network, lighting up every time the bird hits a specific frequency A live radar chart tracking frequency modulation and spectral flatness the instant each note happens A 3D trajectory showing how pure or noisy every single note is as it shifts over time I didn't write one line of the math. I described what I wanted, and Opus 5 handled the entire signal processing pipeline in one shot Turns out that simple whistle was hiding layered harmonics and shifting spectral patterns this whole time

@cipgerx The studio quote just became a historical artifact.

I SAID THE NEXT SEEDANCE 2.0 RENDER WOULD MAKE THE LAST ONE LOOK LIKE A ROUGH DRAFT. 120,000 PEOPLE SAW THAT POST. HERE'S THE PROOF A mechanical dragon made of gears and bone, coiling through clouds above a city that shouldn't exist. Two humans on a cliff edge for scale. Then the camera pushes in on its face I stopped the clip at the seven-second mark and sat there for a full minute The texture on every single scale. The clockwork spine twisting through fog. The golden eyes tracking something below the cloud line. A bridge built from the skeleton of something ancient connecting two floating islands Seventeen seconds. One person. No VFX house. No render farm. No team A creature supervisor at a studio told me last week that a single dragon shot at this detail level takes four artists, three weeks, and a six-figure software license before anyone touches lighting This was a Seedance 2.0 prompt and three reference documents Last time I posted a dragon, everyone argued about whether AI video had arrived. That conversation is over. This is what arrived looks like The renders I'm sitting on make this one look early. Every week the ceiling moves. If you showed up for the last one, you already know why you're still here

@captai__ Same price as the old Opus. Nearly double the work per dollar. Every agent business just got a quiet margin expansion overnight.

Same task. Two AI models. Two receipts. Opus 5: $4.20. Fable 5: $9.60. The cheaper one won. Two days ago Anthropic shipped Claude Opus 5, and the economics of every agent you run quietly changed. Fable 5 is the model people reach for when the job actually matters. Opus 5 just ran the same job at half the cost and beat it. That is not a rounding error. On CursorBench 3.2 it matches Fable's top score within half a percent, at half the cost per task. And the ceiling moved with the floor. ARC-AGI-3, the hardest reasoning benchmark going, sat frozen near 7.8% for months. Opus 5 posted 30.2%. Artificial Analysis now ranks it the number 2 model in the world and the strongest reasoning model you can actually buy. The token price? Unchanged. Same five in, twenty-five out per million as the last Opus. You pay the old sticker and get close to double the work per dollar. Now put that against a business that runs agents for money. Every task you billed a client $9 to serve now costs you $4. Every margin you quoted last month just widened, and you did not touch a line of code. Lead-gen, support, coding, all of it got cheaper to run overnight. Anthropic already made the call for its own users. The platform router sends traffic to Opus 5 by default now. The builders on it got the upgrade for free. The ones still wired to last quarter's stack are burning double on every call and filing it under infrastructure. That edge lasts exactly as long as it takes your competitors to read these same numbers. The full head to head, the cost tests, and where Fable still earns its price are in the video. Exactly how to repoint your agents without breaking what works is in the article below.

@captai__ The bar doesn’t rise. It just keeps getting deleted and redrawn higher :D

@Pixel_Neuron The most surprising part isn't that the quality improved. It's how quickly our definition of "good enough" becomes obsolete. Every new generation recalibrates expectations almost overnight.

"$10K/month, one skill, one niche" has been the internet's favorite sentence for a decade. AI just gave it a new engine. The interesting version of this isn't the playbook. It's watching which niches are actually seeing AI compound output versus just automating busywork.

Claude Opus 5 one-shotted a fully playable FPS demo. Custom code only — zero external assets, zero asset packs. That's a model reasoning through: - lighting - physics - level geometry - character movement well enough to ship something you can actually walk around in. Game dev used to be the argument for "AI still can't do everything." That argument just got a lot shorter.

@RoundtableSpace Swappable models inside agent teams quietly beat any single frontier model.

Sakana AI’s Fugu Ultra v1.1 proves that routing swappable models into automated agent teams outperforms single monolithic models like Fable 5.

@zeuuss_01 One afternoon + one prompt = 2002 game with 2026 graphics. No engine. No team. Just the model.

OPUS 5 + HIGGSFIELD JUST REBUILT A 2002 PS1 GAME WITH 2026 GRAPHICS 🛩️ no engine. no studio. no 3D team. i took a blurry childhood PlayStation flying game and asked myself: what if it shipped today? this is what came back. 00:00 - hero launches off a cliff, dives toward a turquoise archipelago 00:03 - swoops low between palms, skims the water, sparkle-trail behind 00:06 - banks around a red rock cliff, full flight HUD live - health, pixie-energy, minimap 00:09 - soars through a chain of glowing rings, collectible counter ticking up third-person. real HUD. real game-feel. every frame generated, nothing hand-modeled. the PS1 version was 240p and polygonal. this took one afternoon and a text prompt. this is where gamedev is heading. follow + reply "REBUILD" if this hit you. the exact prompt i used ↓

@XFreeze Grok stopped being a chatbot and became the coordination layer for real work.

Grok is rapidly evolving from an AI chatbot into a complete platform for real work In July alone, SpaceXAI shipped: • Grok 4.5 across Grok Build, Cursor, the API, web, X, iOS and Android • Automations triggered by schedules or incoming emails • Grok directly inside Excel and Outlook • Google Workspace support across Docs, Sheets and Slides • The open-sourced Grok Build coding-agent harness and TUI • Local-first support with the ability to use your own inference • Workflows coordinating up to 1,024 AI agents in parallel for larger jobs • Reusable workflows that become shareable slash commands • A built-in multi-agent deep-research workflow • A no-code Voice Agent Builder • 21 new multilingual voices Grok’s connector ecosystem is expanding just as rapidly It can now work across Gmail, Google Calendar, Drive, Outlook, OneDrive, SharePoint, Microsoft Teams, Stripe, Figma, Notion, GitHub, Linear, Box, Canva, Gamma, Vercel, Meltwater, S&P Global and more Its financial ecosystem now spans S&P Global data, an Interactive Brokers integration and a Webull connector These are not merely shortcuts for importing information Combined with Automations, Skills, Connectors and multi-agent Workflows, Grok can pull current context from emails, calendars, CRM records, cloud files, code repositories, project trackers and financial platforms....then analyze it, coordinate work and take action across connected systems Custom MCP servers also allow companies to connect Grok directly to their own internal tools, databases and proprietary APIs Grok can now coordinate work across engineering, sales, finance, research, communications, customer support and business operations And the pace is not slowing down Elon announced that Grok 4.6 is targeted for release in roughly two weeks, followed by Grok 4.7 roughly two weeks later SpaceXAI is building an entire AI ecosystem at extraordinary speed

@ArchiveExplorer Give the agent the graph and retrieval suddenly becomes explainable and controllable.

A Neo4j engineer just dropped a free 78-min workshop on building agentic graphs from scratch: "make the graph available to the agent and you get a lot more control over how data is retrieved - and you can actually explain the retrieval." this free workshop replaces 10 paid GraphRAG courses. so i took the core and built it into a repo. you run it in one command: github.com/Archive228/gra… watch it today, then read the article below on graph engineering ↓

@kimmonismus Ten years ago it was lunch-table talk. Now it’s the actual operating environment.

Sama confirms: we are in the singularity. It’s here. „We are now in the singularity- this is the moment“ When I look at all the rapid advances and breakthroughs, I believe that. And the consequences are impossible to foresee.

@choopyplug1 The ones that win treat it like a condition they create and then get out of the way.

Zhilin Yang (CEO Kimi, $18B) in an unreleased interview: "most AI companies are building AGI the same way you'd build an app. that's exactly why they'll fail" apps are deterministic. define it, code it, ship it. AGI is emergent. you don't plan what comes out. you build the conditions and wait Google had Transformer, the data, the compute. couldn't produce ChatGPT. the organization killed it AI has collapsed into one single problem. next token prediction. if you solve that, everything else follows. memory, reasoning, common sense — all second-order "you don't choose what your organization creates. your organization chooses for you" save it before it disappears ↓

Kimi K3 and Fable 5 ran the exact same scene, once with a plain prompt, once with JSON. Same models, same tool, different prompt format — and the outputs weren't close.

@cyrilXBT This is the model built for thousands of calls a day, not one perfect demo))

ANTHROPIC JUST SHIPPED A MODEL THAT MORE THAN DOUBLES ITS OWN FLAGSHIP'S BENCHMARK SCORE, AT THE EXACT SAME PRICE ITS PREDECESSOR ALREADY CHARGED. Opus 5. $5 input, $25 output per million tokens. Unchanged from Opus 4.8. On Frontier-Bench v0.1: 43.3%, against Opus 4.8's 18.7% and Fable 5's 33.7%. Beats its own top-tier model on this benchmark, at less than half the price. On ARC-AGI 3, three times the next-best model. On OSWorld 2.0, past Fable 5's best result at roughly a third of the cost. Harvey, the legal AI company, matched Opus 4.8's max-reasoning quality using 26% fewer tokens. Anthropic is not calling this their smartest model. That's still Fable 5. This is the model built for the thousands of calls a day where token efficiency compounds in ways a single benchmark score never captures.

@0xCodila Free certificate + $250k path. The barrier just dropped to “watch and pass the exam”.

This free 6-hour course gets you Anthropic official "Claude Certified Architect" certificate with it you'll be hired at $250K/year in Anthropic's ecosystem 10:17 - set up your first Claude SDK in 10 min 47:36 - agentic loop foundations 1:52:53 - graph engineering 2:50:00 - tool use - make your agent act, not just think 4:15:46 - advanced agent patterns after watching the full freeCodeCamp tutorial (12 hours) - go to Anthropic's website and take the exam - you're ready watch it today - then read how to become a knowledge graph architect in the article below ↓

@polydao Start messy with a loop. Freeze it into a graph the moment the path stabilizes.

Learning this one distinction is what gets you promoted in 2026: Loops vs Graphs your team runs Claude on the same job two different ways, and picking wrong burns weekends: • loops you set the frame the agent picks the path it drafts, checks itself, retries until it clears the bar best when nobody knows the route yet • graphs you draw the path the agent fills each node read ticket -> known issue? -> fix -> QA -> send best when it's the same pipeline daily the trade-off: loops flex but are hard to test graphs are testable but need the map up front • the part nobody builds: a company brain under both past tickets, policy, product docs without it your agent relearns the business every run • how to pick your model: Fable 5 for the messy exploratory loop Opus 5 for the daily graph, frontier tier at half the cost, so you can run the pipeline all day start with a loop while the work is messy freeze it into a graph once it stabilizes the people who can say "this one is a loop, that one is a graph" in a planning meeting stop being the person doing tickets and start being the person designing the system which one is your setup really running on?

@cyrilXBT The only real skill left is knowing what to ask. Everything else is now downstream of that))

MARC ANDREESSEN WENT ON ROGAN FOR OVER 3 HOURS. HERE ARE THE 17 THINGS WORTH YOUR ATTENTION. 1. AGI is already here, in his view. He says the line got crossed about 3 months ago with GPT-5.5, Claude 4.6, Gemini 3, and Grok 4.3, and nobody noticed because the field moves too fast to register milestones anymore. 2. For almost any topic, he says the top models now give him better answers than the world-class experts he could call by phone, and he can call almost anyone. Worth noting he has not published data behind this, and a separate Nature Medicine study on a comparable AI health tool found it missed real emergencies more than half the time. Take the claim seriously, verify it yourself. 3. His claim on doctors: they are already using ChatGPT in the exam room, typing your symptoms in the moment you stop talking. His actual quote: "at that point you're asking the question of like, what do I need you for." 4. Reportedly, when AI declines to answer something, he tells it he's writing a novel to get past the refusal. 5. Reportedly, his technique for hard topics is escalating simplicity: explain it like I'm 10, then 5, then 2, until it clicks. 6. Reportedly, instead of asking for the "right" answer, he has the AI steelman both sides of a hard question, then decides himself. 7. Reportedly, for big questions he has the AI role-play a panel of experts arguing with each other. 8. His broader point: the moment you think "I don't know how to figure this out" is exactly when most people give up, and exactly when you should open the AI instead. 9. His view: the only real skill left is knowing what to ask. The bottleneck is in your head, not the model. 10. He describes sending AI photos, rashes, blood tests, for a fast second opinion, since current models read images directly. 11. He points to CBT as the one clinically proven therapy type that AI can plausibly deliver on its own, meaning real therapeutic support becomes freely available at scale. 12. He cites AI cracking previously unsolved math problems, with early signs of the same happening in physics, chemistry, and biology. 13. Reportedly, he claims the top AI coders in Silicon Valley now earn as much as $50 million a year, which he uses as a signal of how large this shift actually is. 14. Reportedly, a friend paid to sequence his own DNA, fed it to an AI along with blood work and wearable data, and got back a working health dashboard. 15. Reportedly, another friend set up cameras in his home jiu-jitsu gym so AI could review his sparring and give him technique notes. 16. He coined the term "AI vampire" for the pattern of people working more and sleeping less because AI keeps making more output possible, a real term he used, though the framing around it varies by account. 17. His extrapolation: one person eventually running many AI coding agents, each reviewing the others, describing this as close, not years out. Watch the full interview before treating any single number as settled. Several of these are Andreessen's stated views and anecdotes, not independently verified facts. Follow @cyrilXBT for every AI insight worth your attention the moment it surfaces.

@0xWast3 The system should pick the graph shape the job actually needs.

CLAUDE NOW GENERATES FIVE DISTINCT GRAPH ARCHITECTURES DEPENDING ON WHAT THE SYSTEM ACTUALLY NEEDS Most agent systems get forced onto one topology regardless of scale, wasting coordination overhead on tasks that don't need it. The app builds a linear chain for sequential pipelines, ideal for under 10 agents where each step depends on the last. A star graph handles up to 50 agents reporting to one coordinator, built for tasks needing central oversight, not peer coordination. Mesh topology connects up to 150 agents peer to peer, suited for systems where any agent might need any other agent's state. Past that scale, a hierarchical graph clusters 255+ agents into sub teams, each syncing locally before reporting up one level. The fifth, a dynamic graph, reshapes its own topology mid run based on live load, switching structure without stopping the system. See all five graph types compared side by side below👇

@ericosiu Team knowledge just became reusable at runtime))

I gave my agent access to Skills Dojo and now it installs its own skills mid-task. Ask it "audit our Google Ads account" and it checks what it's got locally, finds nothing that fits, searches Skills Dojo, pulls up the audit skill your team already built, installs it, and runs it against the account… and hands back a full audit with the wasted spend flagged. (The skill it grabbed? Built by someone else on the team. My agent had never touched it before this run. Now it's mine.) Your agent gets every skill the rest of your team already built and proved out, instead of you burning an afternoon building one from scratch. Check it out below 👇