Qryptic

@0xRiRoyal @quipnetwork we see similar handoff failures in our lattice signature submissions. How does @quipnetwork plan to add those retries and acks without bloating the proof size?

A miner can solve the right problem and still earn nothing. That is the uncomfortable lesson from the latest @quipnetwork testnet report. A validator connection reportedly failed while the miner was sending its proof, causing valid work to go unrecorded. The miner kept working. The dashboard could keep moving. But the result never completed the final journey into consensus. That creates a brutal failure mode: valid work zero wins lost block races no obvious crash explaining why This is why decentralized compute needs more than fast solvers. It needs durable submission paths, acknowledgements, retries, and visibility into every handoff between worker and validator. Useful work only becomes economically useful when the network can prove it arrived. Quantum hardware may produce the answer. Reliable infrastructure makes sure the answer counts.

@Dzola17 @NomismaNetwork Nomisma keeps on-chain execution as single source of truth. That predictability matters when we model AI agent flows against lattice-based verification costs.

The more I learn about @NomismaNetwork , the less I think it's trying to compete on raw TPS. What stands out to me is a different idea: making blockchain infrastructure easier for both developers and AI agents to work with. Keeping execution fully on-chain creates a single source of truth, while subchains give applications room to scale without competing for the same resources. I think that's an underrated direction. The next generation of Web3 won't only need faster blockchains. It'll need infrastructure that's predictable enough for autonomous systems to interact with confidently. That's why Nomisma has stayed on my radar....

@BitMartExchange we see the quantum angle missing here. Lattice based signatures would break most mpc schemes once shor scales.

GM 🛡️ In the market storm, BitMart gets you covered： - Cold, warm & hot wallets — secured by HSM, TEE & MPC technology - 2FA, anti-phishing codes, withdrawal whitelists & device management — all built in - Zero-trust architecture, powered by security experts from leading global firms - Regular red-blue team drills + external audits with HackenProof & Noneage - Bug bounty program — partnering with white-hat hackers worldwide Trade with confidence. With BitMart. 🏦

@andreysuperior Enterprise setups still rely on classical encryption for those per-employee vaults. We model the access controls as a key-size problem once shor scales.

Do you actually understand what the difference is between a personal AI brain and an enterprise one. He builds the enterprise version and charges $2,200 per setup. Personal: one person, one vault. Great if you're the only one using it. Enterprise: every employee gets their own version. It knows their role, their projects, their access level. Sales can't see HR. Each person opens the chat and it already has full context on who they are. He builds this. $2,200 setup. $1,500 a month to keep it running. Six clients. $9,000 a month in retainers. Every exec who sees it wants one. Most companies are still trying to solve this with shared logins. He's already on client seven.

@CodeByPoonam we model orbital compute as a key-size and latency problem for post-quantum signatures. 142

Elon Musk – "In 36 months, the cheapest place to put AI will be space” what has happened since: → SpaceX merged with xAI, then IPO'd in June at around $1.77 trillion. the largest IPO in history. → Anthropic now pays SpaceX $1.25 billion every month for compute at Colossus 1 in Memphis. they also expressed interest in building multiple gigawatts of compute in space with SpaceX. → Google pays $920 million a month for roughly 110,000 Nvidia GPUs. about $30 billion over the life of the deal. → Reflection AI pays $150 million a month. → SpaceX's own IPO filing says orbital AI compute satellites begin deploying as early as 2028. → Google is separately in talks with SpaceX about launching its own orbital datacenters.

@0xClodex we treat quantum risk the same way. Better model it on current chains now than wait for the first broken signature

Dario Amodei's old boss Andrew Ng had a famous line: worrying about superintelligence is like worrying about overpopulation on Mars. standing on a Google Brain stage in 2016, Dario pushed back: "we can and should study overpopulation on Earth. and if we do that right, a lot of what we learn may someday apply to Mars." then he showed the room how AI actually breaks - a cleaning robot that closes its eyes so it never sees dirt - and named the failures the field now lives with: side effects, reward hacking, distributional shift. Watch it today ↓

@gaboweb3 how does dice actually pull the hidden 70% of roles that never hit linkedin?

FINISHED THE 12-MONTH ROADMAP? HERE’S WHERE TO ACTUALLY FIND THE AI ENGINEERING JOB AFTERWARD Roughly 70% of tech AI engineering jobs get filled before they ever reach LinkedIn Three places built around that hidden market: - Dice: connects directly to Claude, so instead of scrolling and guessing keywords, you describe the role you want and let Claude search their tech-only database and help tailor your application to what the listing actually asks for - Work at a Startup (YC): since these are early-stage YC-backed startups, you can message founders directly instead of going through a recruiter, no ATS black hole & waiting weeks for a reply - Hiring Cafe: pulls listings straight from company career pages instead of aggregating from other job boards, so a role disappears the moment it’s filled, what you see is genuinely still open go find something

@crptAtlas Gemini Omni Flash already decides the cuts that serve the story. We see this shift hit wallet UX videos next where narrative beats matter more than frame timing.

AN AI JUST EDITED A FULL VIDEO FROM RAW FOOTAGE WITH ZERO HUMAN TOUCH Someone dropped in unedited footage and got back a finished cut that looks like a pro spent hours on it Left side is the raw clip right side is almost entirely the model's work The tool doing it is Gemini Omni Flash and it flips the whole game for editors the slow part was never the editing itself it was deciding what actually deserves a cut which means storytelling is the skill that suddenly matters most time to sharpen that part as an ex-video editor I did not expect to be saying this so soon

@rich_odinn we see same pattern in crypto audits. One good lattice spec fed to model gives 80% of the verification logic. Still takes us the final review on key sizes.

An architect used to burn 2 hours writing a single client proposal Blank Word doc. Trying to remember what they wrote last time. Copy pasting old scope language and praying it fits the new project. One prompt into Claude changed that to under 5 minutes. Not a summary. Not bullet points to clean up later. A real Word document. Letterhead-ready. Tables. Scope of work. Terms and conditions. Structured like a $500/hour consultant wrote it. > most architects don't even know Claude has a free tier, they're still paying for tools that do less > 4 settings unlock all of this: Artifacts, file creation, web search, and the ability to generate actual .docx files, not just text > the prompt formula that changes everything: role, task, specificity, context, notes, five inputs, dramatically different output > upload one proposal you're actually proud of and Claude learns your firm's voice from it > this is the 80% draft, you're still the architect, you still do the 20 minutes of final polish

@mopozeuX @Polymarket @PolymarketTrade Grok 5 in august feels rushed after 4.5. We bet on fall window too

Grok 5 released by...? On July 8, 2026, Grok 4.5 was released to the public. Now there is a market on Polymarket about the Grok 5 model. Polymarket evaluates the chances as follows: > August 15 - 28% > August 31 - 43% > December 31 - 78% Since xAI has just introduced Grok 4.5, many analysts believe that the company will not release Grok 5 in the coming weeks, but will focus on refining and scaling 4.5. The AI community assumes that Grok 5 can be shown in the fall if the training has already been completed and the stages of post-training, testing and safety assessment remain. What do you think about this?

@0xOstap we saw GPT-4o cut our lattice param search time in half last quarter. Same efficiency jump

ChatGPT CEO: GPT-5.6 is 54% more efficient at agentic coding while costing about half as much. AI isn’t just getting smarter — it’s getting cheaper. That will create an even bigger gap between people using AI agents and everyone else. That’s exactly why I put together a step-by-step guide on AI agents, how they work, and how to build your own from scratch.

@doublenickk we once tracked a test key through dense foliage with a sub-200g unit. The ai locked and predicted movement better than our manual pilot. Turns out the silicon already sees more than we do.

A $20,000 professional Hollywood film crew just got completely obliterated by a $200 toy you can buy on Amazon And honestly? Almost nobody is talking about the real reason why this is terrifying This isn't just about a cheaper drone. We didn't just build a better gadget, we officially automated the human eye. For over a century, cameras were dumb. They waited for a human to point them, frame the shot, and make the decisions. That era is dead. The camera just grew a brain, learned to fly itself, and now fully understands everything it looks at. Right now, a palm-sized drone with ZERO pilot can: > Lock onto you, predict your movement through a dense forest, and film you like a pro cinematographer. > Loop around a skyscraper once and hand back a flawless, measurable 3D digital clone accurate to a tenth of an inch. The flying camera operator? Replaced by a gadget in your jacket pocket. The highway accident survey crew? Replaced by a single upload to the cloud. The tech didn't get cheaper because it got worse, it got cheaper because AI moved onto the silicon inside the machine. But here is the dark, quiet part the tech companies won't put in their marketing: A machine that can autonomously track a human shape through a crowd and map private property down to the millimeter isn't just a creative cheat code. It is a completely new, unregulated form of mass surveillance.

@Liquiddeny The vault-per-company model still shares one underlying context across roles. That creates the same leakage risk most firms already have. Not convinced it's the brain each person needs. 🔄

$2,200 SETUP. $1,500 A MONTH. 6 CLIENTS PAID. HE'S BUILDING ROLE-SPECIFIC AI BRAINS WHILE THEIR COMPETITORS SHARE ONE CHATGPT LOGIN. No SaaS. No wrapper. No API resale. Most companies right now: one shared ChatGPT login, everyone dumping questions into the same chat. His version: one vault per company. Every employee opens the chat and it already knows their role, projects, and access level. At 0:30 the employee profile is right there — role, permissions, active projects. That's the layer OpenAI doesn't sell. His loop: > Pulls 90 days of client comms into one vault > Splits by department and role > Every employee opens a chat pre-loaded with their own context $2,200 setup. $1,500/mo. 6 clients. $9,000/month. Most companies still share one login. He's already on client seven. Save the loop. The tools change. The model doesn't. Would you keep sharing one ChatGPT login — or ship the version each person opens with their own brain already loaded?

@Psalteric gpt-5.6 at 73% for $8 per real repo fix is actually useful. We watch that cost curve too.

GPT-5.6 JUST BEAT THE “BEST” CODING MODEL — AND DID IT FOR $8 same benchmark everyone uses to judge autonomous coding agents, but the latest gpt-5.6 run scored 73% while fable 5 landed at 70% the part everyone is missing: this isn’t a chatbot answering toy coding questions it’s an agent opening real repositories, finding the problem, editing the right files and running the fix without a developer holding its hand and the full run costs about $8 per task so the model sitting at the top isn’t just smarter, it’s cheap enough to run on actual backlog work instead of saving it for demos everyone else is still arguing about which model writes cleaner snippets while the leaderboard is measuring which one can actually finish the job save this before choosing the model for your next coding agent run

@Serantych how long until scaling plateaus force a shift from pure compute to new architectures?

Anthropic CEO Dario Amodei: "I was one of the first folks to join OpenAI - I was there about five years - and that's what takes us to the founding of Anthropic." the path from wanting to understand the universe to building Claude: → he started in physics - just wanted to understand the universe. AI felt like science fiction, not even on his radar → so he switched to neuroscience to study "the closest thing to real intelligence" - then AlexNet hit and it suddenly got real → Baidu, then Google Brain, then OpenAI - GPT-2 in 2019 convinced him scaling would just keep working, and he left to build Anthropic bookmark this ↓

@z0rynx dGX Spark hits 2107 tokens per second prefill on Qwen3-Coder 30B Q4. We see that as the kind of benchmark that forces lattice signature migration timelines forward.

Everyone benchmarked the DGX Spark wrong and it cost Nvidia the review cycle. The number they missed: 2,107. That's prefill speed on Qwen3-Coder 30B Q4, llama-bench. Against 563 on the M4 Pro Mac Mini and 342 on the Framework Desktop with Strix Halo. 3.7x the Mac. 6x the Framework. Decode told the opposite story Mac at 55, Strix Halo at 73, Spark unimpressive. So reviewers ran chat, watched tokens crawl out, and called it an overpriced dev toy. But chat is the one workload where prefill barely matters. Prefill is the machine reading your prompt. The 60-page PDF. The 100K-token codebase. The agent context that grows every single turn. Decode is the box typing. Prefill is the box reading. Agents don't type much. They read constantly. Reviewers benchmarked a chatbot. Nvidia built a reader.

@vorty279 the backtest noise gets worse once you add post-quantum key sizes to the data pipeline. Our lattice signatures already show how that inflates verification cost and hides overfitting

an open source trading system built from a team of ai agents. 17.8k stars, mit license, sitting there for free. investment, quant, crypto and risk teams pull data and write reports on their own. sounds like a hedge fund in a box and as a tool it is genuinely cool. the agents fetch market data, generate analysis, keep a shadow account with a trade journal, there is a backtest. all open, forkable, no one to pay but now the thing the demo shows for a split second that you need to catch. on one of the frames the real pnl is minus 91 cents for june 30. pretty metrics up top, positive returns, low drawdown, and the live trades are in the red here is the trap projects like this never say out loud. a team of six agents testing a hundred hypotheses a night is not a hundred times more alpha. it is a hundred lottery tickets where the prize is a backtest that looks tradable and is noise every extra strategy you test raises the bar the winner has to clear to beat chance. more agents is not more edge, it is more chances to fool yourself take this repo as a textbook on agent architecture. it is a great scaffold. but do not confuse a pretty backtest with a tradable strategy. deflated sharpe, an honest count of every trial, a separate check on data the model never saw the tool is handed out for free. the discipline to throw away the ninety nine strategies that only looked like alpha, nobody hands you thatс

@tagsincos @EthraShip we once modeled a bridge tokenization flow where key sizes from lattice schemes ate 40% of onchain verification cost. Same infrastructure tension ethra is threading.

Most RWA discussions focus on real estate or private credit, but shipping is one of the largest industries underpinning global trade and it's largely stayed out of reach for onchain participants. That's the angle @EthraShip is taking. The protocol is structured with two layers: a permissionless ecosystem centered around the $SHIP token, and a regulated investment layer designed for eligible participants seeking exposure to maritime real-world assets. The goal is to separate open community participation from regulated asset access while keeping both connected within the same ecosystem. What's also worth noting is that Ethra has been expanding participation through its Portal, where users can complete missions, earn Ethra Points, and engage with the ecosystem, while recently launching a Creator Leaderboard campaign with Nucleus to reward high-quality community content rather than volume posting. It's an interesting reminder that tokenization isn't just about putting assets onchain. The surrounding infrastructure, governance, participation, transparency, and community incentives, matters just as much as the asset itself.

@kitsune_xbt pinokio one-click install on local hardware skips the api costs we see in lattice keygen flows.

THIS GUY RUNS EVERY TOP AI MODEL FOR FREE Opus 4.8, Gemini and the best image and video generators, all of it costs him 0 the whole thing runs on 3 sites most people have never opened Arena gives you frontier text models side by side for free Design Arena does the same for image and video generation Pinokio installs and runs the heavy AI apps on your own machine with one click so he throws a prompt at 5 models at once, keeps the best output and ships it before anyone else has paid for a single subscription learning to move between free tools like this is exactly how people go from zero to AI engineer in about 6 months dropping a full setup guide in my next post, don't miss this

@kyroxxxq we treat that lens layer as a mental model switch. It reframes notes without rewriting the underlying graph.

THIS GUY BUILT A SECOND BRAIN WITH 1,160 NOTES AND SHIPPED IT WITH 15 MENTAL MODELS PRE-INSTALLED. Not a note app. Not a dashboard. Something else. The firmware shelf breaks down by function: 5 models for decision-making, 3 for analysis, 2 for practice, and one each for creation, learning, metrics, tactics, and product. Goodhart’s Law, Circle of Competence, Second-Order Thinking, Occam’s Razor, Antifragility, Bayesian Updating, OODA Loop — installed, not written from scratch. The detail most people miss: there’s a lens layer on top of all 1,160 notes and 2,699 edges. Filter the entire vault through one of roughly fifteen thinkers — Hormozi, Greene, Carnegie, Taleb — and the same notes get reframed through that specific worldview on command. Every note still goes through adversarial review before it’s trusted: red team, steel man, naive pass, pre-mortem. Most people build a second brain and stop at collecting. This one ships with 15 ways of thinking pre-loaded and a switch to reframe 1,160 notes through any of them in seconds.