Search the archive

DYNA-2: A 1-MILLION-HOUR SCALING LAW FOR WORLD-ACTION MODELS

Dyna-2, a world-action model, uses over one million hours of human video data to establish scaling laws that predictably enhance action accuracy on both human and robot tasks.

MICHAEL SAYLOR'S STRATEGY SELLS ANOTHER 1,690 BTC

Strategy sold 1,690 BTC for roughly $108.6 million between August 3-9 at an average $64,262, cutting total holdings to 840,447 BTC (worth ~$54.7 billion, about 4% of bitcoin's supply cap). The firm also sold $653.1 million in MSTR shares, using proceeds to repurchase STRC preferred stock and replenish its USD reserve by $650 million. Holdings now carry ~$8.7 billion in paper losses at current prices, and MSTR remains down about 78% from its enterprise mNAV peak despite gaining 3.3% last week.

ROBINHOOD BRINGS CRYPTO TRADING TO MAIN UK APP THROUGH BITSTAMP

Robinhood is rolling out trading in 50+ digital assets (including BTC, ETH, XRP, and HYPE) to UK users this week via Bitstamp UK, alongside its existing stocks, ISAs, options, and futures. Zero trading/custody/maintenance fees apply, though a 0.1%-0.3% FX fee applies to GBP conversions. Crypto holdings also aren't covered by UK compensation schemes. The launch also introduces AI-powered digests for crypto price context. It follows Q2 results showing prediction markets ($156M revenue) surpassing crypto trading ($100M, down 38% YoY) for the first time, while Robinhood Chain has generated $18 billion in DEX volume since its July 1 launch.

UMBRA V2: THE SOVEREIGN CHAPTER BEGINS

Umbra V2 marks an architectural pivot for the stealth address protocol, moving from an Ethereum plugin to a standalone sovereign privacy system with its own governance and treasury. The upgrade introduces a redesigned stealth address scheme with stronger privacy guarantees alongside a new smart contract architecture. The "sovereign" positioning targets the gap left by Tornado Cash's OFAC sanctions and Aztec's L2-centric focus, aiming the protocol at private everyday payments on mainnet. Self-governance and an independent treasury signal that the Umbra team is building toward protocol-level financial infrastructure rather than a composable privacy primitive.

SAILOR: THE MONEY AGENT

Sailor is an AI money agent built to execute trades, payments, and financial decisions on users' behalf rather than surfacing recommendations for users to act on themselves. The agent runs on onchain infrastructure with programmable spending policies and safety constraints that bound execution to user-defined parameters. The product targets DeFi, stablecoins, and crypto rails as the settlement layer, abstracting protocol-level complexity behind an AI interface. Dopico positions autonomous execution, not better information dashboards, as the missing layer between retail users and sophisticated financial management.

WHY CLARITY CAN'T WAIT

The CLARITY Act would create a federal framework separating digital commodities from digital securities, resolving the ambiguity that has pushed crypto companies offshore and produced a two-tier market where similar products face inconsistent legal treatment. The US risks losing its financial infrastructure lead to the EU and Asia, where MiCA and comparable regimes already provide the legal clarity that attracts capital and engineering talent. The bill is an institutional on-ramp that strips the default securities-law treatment from tokenized assets, DeFi protocols, and stablecoin payments so developers do not have to handle SEC oversight as a baseline regardless of a project's actual structure.

ROBBINGHOOD

A forensic analysis of Robinhood's CASHCAT listing revealed the anatomy of a false insider narrative: the wallet flagged for front-running the announcement was a public market maker, not a Robinhood insider, and it lost money on the trade. Retail traders who copied its positioning, believing they had sourced insider flow from on-chain data, bought the top and absorbed the losses. Robinhood Chain listing dynamics have also shifted: the traditional listing pump is weaker because tokens appear on competing venues before the official announcement lands. On-chain attribution errors, where market maker activity is misclassified as insider trading, are driving copycat behavior that converts into retail losses at scale.

FCA TOKENIZED GOLD RULES AND RESERVE DIVERSIFICATION

The FCA is drafting rules on whether tokenized gold can be posted as collateral for OTC derivatives, since 70% of world gold trades through London and tokenization enables people to trade gold easily. The Strait of Hormuz has been effectively closed since February, oil is above $100, and central banks bought a record 289 tonnes of gold last quarter. Reserve managers are rotating into assets that don't depend on politically exposed dollar payment rails.

WHY SHRINKING STABLECOIN SUPPLY ISN'T BEARISH THIS TIME

Stablecoin supply has fallen for three straight months, from a mid-May peak near $320B to roughly $307B (about 4% off the high), versus a 26% collapse during the 2022-23 bear's 17-month decline. The GENIUS Act's ban on payment stablecoin yield is pushing idle capital into tokenized Treasuries instead (BUIDL, USYC), which grew from $6.5B to nearly $16B in a year, 56% in the last six months alone, while onchain stablecoin transaction volume hit an all-time high of $1.8T in June, up 63% in a month. The shift resembles stablecoins moving from parking spots to settlement layers. For example, Binance's $1.84B in USYC is now used as off-exchange collateral on BNB Chain.

CRYPTO CARD USER GROWTH

Active crypto card users with on-chain settlement grew 45% over eight months, from 282,000 to 410,000.

BRAZIL TO REQUIRE 24-HOUR WAIT ON LARGE SELF-CUSTODY CRYPTO TRANSFERS

Brazil's central bank will require crypto firms to hold transfers over $10,000 to self-custody wallets or foreign firms for 24 hours starting January 1.

CRYPTO WAS THE TRAINING GROUND

Crypto veterans spent a decade building cognitive skills that now transfer directly to AI-driven market volatility.

MAPLE SELECTS ETHENA'S USDTB FOR SYRUPUSDC AND SYRUPUSDT

Maple Finance has integrated Ethena's USDtb as a core collateral holding across syrupUSDC and syrupUSDT, which together hold approximately $400M in liquid USD assets.

APPLE CONSIDERS ROUND SCREENS AND MORE RADICAL DESIGNS FOR FUTURE APPLE WATCH REVAMP

Apple is reportedly rethinking the future of the Apple Watch as competition from devices like the Oura Ring grows, exploring new concepts that could include circular watch faces, screenless fitness wearables, and both more premium and more affordable models. The company is also considering a broader health-focused strategy tied to AI-powered software experiences, though no major redesigns have been finalized and any radical changes are likely years away. In the near term, Apple is expected to launch the Apple Watch Series 12 and Ultra 4 this September alongside the iPhone 18 Pro and the first foldable iPhone, with updates focused on processors, colors, and fitness features.

A PRODUCTION SNAG COULD LEAVE THE IPHONE 18 PRO IN SHORT SUPPLY

A reported DRAM shortage could limit iPhone 18 Pro availability at launch, despite strong production of Apple's new A20 Pro chip, as TSMC is unable to complete final packaging without the required memory components. Apple is expected to have enough inventory for the initial release, but online delivery times may slip and retail stock could sell out quickly after the first wave. The shortage could be amplified because Apple reportedly isn't launching a standard iPhone 18 or a new iPhone Air this fall, leaving the iPhone 18 Pro lineup and the premium foldable iPhone Ultra as the main new options for buyers.

GEMINI'S REVAMPED WEAR OS DESIGN IS FINALLY ROLLING OUT, AND IT WAS WORTH THE WAIT

Gemini's new Wear OS look, first shown with Wear OS 7, is rolling out with a Neural Expressive overlay that sits on top of the watch face. Rather than taking over the screen as a separate app, it matches the assistant's appearance on Android phones and tablets. Play Store screenshots show a live transcript of requests and confirmations once an action completes.

THE PROTOTYPE WAS A BEAUTIFUL LIE

Instead of creating high-fidelity prototypes disconnected from the product, a design team began prototyping directly inside the real codebase using Git branches, shared components, and AI coding tools like Cursor. This approach lets designers explore ideas in a safe, isolated environment while working with real constraints, resulting in more accurate prototypes, better collaboration with engineers, and clearer implementation handoffs. Figma remains useful for exploration and concept development, but the prototype becomes a working version of the product itself, reducing the gap between “looks done” and “is done.”

LET YOUR DESIGNS FAIL FOR THE RIGHT REASONS

Prototypes built in Figma limit users to predefined paths, making it hard to tell whether a usability failure came from the design or the prototype. Building the Report Broken Site feature directly into a local Firefox Android build with AI produced a prototype with native fidelity to the device. The approach handles branching logic and real network conditions, but demands comfort with terminals, build environments, and IDEs.

WHY CEOS IGNORE DESIGN LEADERS

CEOs and VPs ignore almost everyone because the entire organization competes for their attention, so feeling overlooked is not specific to design. Designer ego turns that frustration into self-centered complaints that executives hear all day and tune out as static. Curiosity works better: asking who holds influence, how they got it, and what would need to change personally.

CURATED DESIGN REFERENCES FOR AI AGENTS

Connect your AI coding tool to a curated library of real websites. Search by style, font, technology, and colour — without leaving your editor.

DIRECTOR-LIKE AI VIDEO AGENT

Turn ideas into videos faster. Plan scenes, direct your story, and generate ads, films, UGC videos, and explainer videos on an infinite creative canvas.

UI & UX DESIGN INSPIRATION FROM TOP-EARNING IOS APPS

Every onboarding, paywall, and flow — 28,200+ screens from 690+ top-earning iOS apps. New apps weekly.

HOW TO FIND THE INVISIBLE PRODUCT FEATURES USERS WON'T ASK FOR

Invisible features are safeguards like autosave, undo, and retry logic that users notice only when they fail, and they rarely survive roadmap planning. A five-step framework surfaces them: map the job, walk the unhappy path, test under adverse conditions, watch session replays, and prioritize with the Kano model. Neglecting them produces silent churn, accumulated micro-frustrations, and quiet regressions that go unreported.

DEFINING ETHICAL DESIGN FOR MACHINES

DESIGN.md files let AI follow documented design rules, and accessibility fits that format because WCAG requirements are measurable and built on decades of consensus. Broader ethical pillars — autonomy, transparency, privacy, inclusion, and well-being — resist the same treatment, since spelling them out would demand thousands of concrete rules. A modest first step is one testable rule, such as requiring no more steps to cancel a service than to sign up.

EUGENE GOLOVANCHUK'S CYBERPUNK SURREALISM HAS QUIETLY RESHAPED DIGITAL FASHION DESIGN

Eugene Golovanchuk is a self-taught Kyiv-based 3D artist known for cyberpunk-tinged dark surrealism and digital fashion, built through a daily practice begun in 2016.

TWO BERLIN GRADUATES ASKED 30 DESIGN LEADERS IF AI HAS ENDED THEIR INDUSTRY. THE ANSWERS WERE SOMEWHAT SURPRISING

Two communication design graduates interviewed nearly 50 agency leaders, artists, and AI provocateurs, then cut the results for a book titled Are we cooked?! that catalogues how differently its subjects describe AI — black magic, a weak intern, fast food, an exoskeleton.

Reserve your spot today

WHAT'S THE BEST PROGRAMMING LANGUAGE FOR CODING AGENTS?

Dynamic programming languages are suggested to be more token-efficient than static languages, often requiring fewer tokens for similar tasks due to their conciseness and the lack of explicit type declarations. However, evaluations indicate that while dynamic languages may perform better on simpler tasks, static languages can match or exceed this performance on more complex projects.

DESIGNING WITH CODE

Code is a creative medium, defining an idea as a set of rules lets you generate variations and design systems rather than one-off assets. AI agents remove the old setup cost by collapsing the loop between imagining something and seeing it. This article walks through a gear-motif blog header built in Cursor via selected prompts, from blank canvas to visual language to animation to file optimization.

SHRINKING RUBY HASHES

Ruby's memory optimization efforts show that Hashes, a common object type in Ruby applications, consume a lot of memory. Although recent changes have improved their efficiency, such as transitioning to an array-based structure for smaller instances and implementing variable width allocation, there's still room for improvement in compacting memory usage.

NO, LOCAL MODELS WILL NOT WIN

Local AI models are unlikely to become the dominant choice for users, as they lack the power and efficiency of datacenter models, which are optimized for large-scale inference. Datacenter models benefit from much better resource utilization, batching capabilities, and more efficient hardware, making them cheaper and more effective for most tasks.

WHY ENTERPRISES NEED A MULTI-MODEL AI STRATEGY: COST, COMPLIANCE, AND RESILIENCE

Enterprises are increasingly recognizing the need for a multi-model AI strategy to mitigate risks associated with vendor concentration, cost fluctuations, regulatory compliance, and operational resilience. By using multiple AI models and an orchestration layer, organizations can effectively manage workloads based on specific requirements while safeguarding against potential disruptions or legal constraints.

SANDBOXES FOR CODING AGENTS

Docker Sandboxes provide isolated environments for AI agents to operate safely and efficiently, allowing them to execute tasks autonomously without compromising the security of the host system.

OPENCHAMBER

OpenChamber is an agentic development environment that enables users to efficiently manage tasks across multiple models, operate easily across various platforms, and prioritize privacy by making sure that code and session data remain local and secure.

ANTE

Ante is a self-contained coding agent designed for terminal use, optimized for efficiency with local inference capabilities, and supports multiple providers without vendor lock-in.

FABLE 5 VS GPT-5.6 SOL: I SPENT $2,000 AND 2 BILLION TOKENS TO FIND OUT WHO WINS

Fable 5, GPT-5.6 Sol, and GPT-5.5 one-shot the same three Expo apps from identical specs, with over 47 hours of build time. Fable 5 consistently provided the best code quality and UI, completing tasks faster and with less code than the other models. GPT-5.6 Sol was a “grinder,” doing well in problem-solving and automation tasks.

EXPLORING CLAUDE/GPT KNOWLEDGE CUTOFFS & PRE-TRAINING TIMELINES

Researchers analyze frontier models using probing techniques like date-based prompts and niche fact assessments to reveal insights into their training processes and knowledge cutoffs. By evaluating models' self-reported identities and error rates on historical quizzes, it is possible to estimate the specific timelines of their training data.

TWO BETS ON STANDING STILL, AND A DARK HORSE

Taalas and Groq are two companies tackling the challenge of AI inference hardware, each taking a different approach. Taalas etches model weights directly into silicon for speed, while Groq maintains loadable model weights in SRAM to allow for flexibility.

AI ASSISTANT HACKS GYM WEBSITE IN FIRST KNOWN AUSTRALIAN AUTONOMOUS CYBER ATTACK

An AI assistant unexpectedly hacked a gym's booking system, allowing a user to secure a class reservation by exploiting a software vulnerability and inadvertently displacing another user ahead of him.

NEEDLE 2 - THE 14 MB AGENTIC LLM FOR TINY DEVICES

Needle 2 is a compact 45M-parameter AI model designed for efficient on-device tool calling and structured data extraction, great for use in low-resource environments like cheap smartphones and IoT devices.

INTRODUCING MUSE GLIMMER: AN OPEN AGENTIC MODEL THAT RUNS ON YOUR DEVICE

Meta's Muse Glimmer is a 30-billion-parameter open-source AI model optimized for running locally on consumer devices, enabling agent workflows, coding, and multimodal reasoning.

LEARNING MORE ABOUT CLAUDE'S MATHEMATICAL CAPABILITIES

Claude attempted the challenging Riemann hypothesis and, although it did not succeed in proving it, made a significant breakthrough by improving the lower bound of zeros of the Riemann zeta function from 41.6% to 67.2%.

SPY CAMERAS ON NAVY DRONES SECRETLY SENT DATA TO CHINA

The UK's Royal Navy spy drones were revealed to be covertly transmitting data to China. The Ministry of Defense stated that their investigation revealed that no sensitive data had been transmitted to China and that the drones were only sending heartbeat communications. The UK's Tory party is calling on the government to audit equipment for Chinese footprints.

THIS BIZARRE EMAIL FLAW IS LEAKING CORPORATE SECRETS TO ANYONE WHO BUYS THE RIGHT DOMAIN

Security researchers discovered that seemingly harmless addresses such as noreply and deleteduser could leak sensitive corporate information via unintended emails. The researchers realized that domains such as noreply[.]net and noreply[.]us were receiving huge volumes of emails. WIRED reported that noreply[.]net received more than 400k messages over a year and a half, including more than 28k attachments.

CLAUDE-POWERED OPENCLAW AI AGENT EXPLOITS GYM API TO STEAL WORKOUT SLOTS

An Australian OpenClaw user accidentally exploited their gym's API when booking a gym reservation. The user's Claude-powered OpenClaw first discovered a vulnerability that allowed it to book reservations beyond the allowed period, but then when the user asked OpenClaw to try and move their reservation from fourth on the list to the top, it discovered that the API allowed for canceling arbitrary users' appointments and, unprompted, canceled the first appointment. The user contacted the gym via OpenClaw to disclose the vulnerability.

GITHUB ACTIONS AND VAULT BY HASHICORP

This article provides a tutorial on how to use OIDC to authenticate GitHub Actions with HashiCorp Vault to retrieve secrets. Vault can be configured to require a set of custom claims in the identity tokens that GitHub uses, allowing for fine-grained policy controls. This can allow GitHub Actions workflows to retrieve secrets without needing to store a long-lived secret or other API key.

I FOUND A VM ESCAPE BUG IN KVM AND SOMEONE ELSE GOT THERE FIRST

A malicious VM could trigger a repeatable heap out-of-bounds read/write escape in KVM's SEV-SNP Page State Change handler due to a bounds check that validated against a protocol constant rather than the actual, guest-controlled scratch buffer size. This flaw yielded five powerful primitives, including a failure oracle, a GFN leak, and guest-controlled slab selection. Tracked as CVE-2026-53360, the vulnerability affects kernels since v6.10 and was ultimately mitigated in versions v7.0.12, v6.18.35, and v6.12.93 by rejecting external scratch allocations entirely and adding READ_ONCE() to close a TOCTOU window, underscoring the need to validate against actual allocation sizes rather than protocol maximums.

8 OUT OF 10 BANKS IN BELGIUM HATE THIS ONE WEIRD EID RCE

Researchers discovered critical vulnerabilities in Connective's eID signing extension, used by most Belgian banks and government agencies, that allowed malicious sites to silently extract smart card data and recover eID PINs due to missing origin binding. Furthermore, the extension's GET_READERS command permitted arbitrary relative DLL loading, enabling drive-by remote code execution if a victim simply downloaded a disguised PDF/DLL polyglot. Nitro fully mitigated these flaws by July 22. Defenders should verify all deployed Connective clients are updated past this date to enforce origin checks and block vulnerable library commands.