Search the archive

INTRODUCING RUN SDK: SECURE EVAL FOR YOUR AGENTS

Run SDK is a package from Vercel for executing untrusted JavaScript or TypeScript without giving it direct access to apps or systems. It evaluates JavaScript or type-stripped TypeScript in a fresh QuickJS context inside a worker thread and exposes selected operations through hostFunctions. Calls cross the sandbox boundary through serialization. A host function may return a promise, so existing service clients can sit behind this interface without being passed into the sandbox.

APPLE HOLDS FAREWELL PARTY FOR TIM COOK AS TERNUS PREPARES TO TAKE OVER

Tim Cook will step down as Apple's CEO on September 1. The company held a celebration for Cook on August 23, a day before his 15th anniversary as Apple CEO. Approximately 200 people attended the party, where the band OneRepublic performed. Cook will remain with Apple and become the company's executive chairman. John Ternus, who has been at Apple since 2001, will take over as CEO.

COULD NVIDIA DEFAULT?

Credit Default Swaps insuring against an Nvidia debt default have doubled in price in the last two months. The company gave a $105 billion residual value guarantee for OpenAI's Ohio data center earlier this month, and it has also been involved in two other $500 billion deals this month. It has been reported to have been involved in over 300 financing deals in the last five years. Analysts are worried an AI downturn could see the circular deals entered into by Nvidia with other AI players unravel, which could result in cascading losses.

SOME RECENT APPLICATIONS OF AI

Some recent applications of AI include gathering mostly unstructured real-world data and checking stuff and high-judgment lifestyle questions.

ENGINEERING THEATRE

Engineering theater is when something looks like engineering but really isn't.

LET ME CLICK

Examples of the 'Let me click' principle in user interface design.

A CAUTIONARY TALE ABOUT DATA BREACH CLAIMS, VERIFICATION, AND CARHARTT

Take headline numbers with a grain of salt unless you're confident in the processes of those making the claims.

X SENDS CEASE-AND-DESIST TO OPEN SOURCE PROJECT NITTER OVER ALLEGED SCRAPING

Nitter and its instances offer a handy way for people to keep up with posts on X without an account.

INTRODUCING INDEX: BUILDING THE WORLD'S LARGEST AND MOST DIVERSE PHYSICAL DATASET

Figure's Index is an app that was created to help the company collect the physical data that its AI stack needs, directly from humans at scale.

OPENAI'S JALAPEÑO INFERENCE ACCELERATOR MOVES TOWARD DEPLOYMENT

OpenAI reported first results from Jalapeño, an inference accelerator designed around low-latency agent workloads, and plans to deploy it in its own infrastructure by year-end. A large connected system keeps prompt processing and token generation close together, while AI helped design circuits and program kernels.

PERPLEXITY PARTNERS WITH NVIDIA TO LAUNCH PORTABLE COMPUTER, A FULLY LOCAL AI AGENT WITH ZERO TOKEN COSTS

Perplexity's Portable Computer is a version of its agentic Computer platform that runs entirely on hardware users already own. The model, user data, and work can all stay on local machines with no billing credits. Every task starts on device by default, and the system asks for permission before sending any individual step to a more powerful model in the cloud. Portable Computer is now available for Pro, Max, Enterprise Pro, and Enterprise Max subscribers on Linux, with Windows support coming in September. Users will need an RTX GPU with at least 24GB of VRAM to run the agent.

VOCAB BREAK

Claude's current tokenizer appears to only have about 15,000 entries. This is surprising, as the trend had seemed to be that more is better in this space. One theory is that Anthropic has been working around a bottleneck caused by the final softmax layer. This article takes a closer look at how Anthropic might be achieving this and the effects it might have on model training.

OPENAI AND ANTHROPIC COULD DOMINATE GLOBAL AI COMPUTE

Dylan Patel discusses how OpenAI and Anthropic could control most usable AI compute by 2028 as their ability to monetize FLOPs lets them outbid competitors. The conversation also covered rising AI capex, potential sovereign debt risks, and the economic forces pushing the industry toward greater centralization.

OPENAI'S JALAPEÑO OPTIMIZES INFERENCE THROUGHPUT AND TOKEN LATENCY

OpenAI's Jalapeño inference chip was designed around its own model workloads, with early benchmarks showing higher peak throughput per kilowatt and lower token latency than the commercial systems tested on

MOATS IN THE AGE OF FLOODS

Abundant frontier intelligence will not eliminate application-layer moats - it shifts value toward companies that translate models into real outcomes. Durable winners will own coordination, workflow data, customer transformation, narrative, higher-level abstractions, outcome-based economics, and structural necessity.

spin.

OPEN OMNIMODAL WORLD MODELS

EchoWM is an omnimodal world model that follows continuous 6-DoF camera trajectories while jointly generating 720p video, environmental sound, music, and speech. It supports first- and third-person interaction and uses progressive plus autoregressive training for synchronized long-horizon generation.

SHORT-LIVED CREDENTIALS FOR AI AGENTS

Vercel Connect replaces long-lived API tokens with runtime-issued credentials that are scoped to individual tasks and expire automatically. Its generally available release added more than 100 connectors along with a unified integration model and production governance controls.

GRANITE 4.2 LLMS: HOW THEY'RE BUILT

Granite 4.2 models by IBM are dense, decoder-only reasoning LLMs available in 3B, 8B, and 30B sizes. Trained on 15T tokens, they use a five-phase strategy that includes a multi-stage RL pipeline and support native tool calling with a THINKING/NON-THINKING switch. The 8B and 30B models learn agentic behavior through RL stages in real environments, enhancing capabilities such as code editing and web searching.

‘THE WORLD SEEMS TO BE READY': AN INTERVIEW WITH OPENAI HEAD OF PRODUCT THIBAULT SOTTIAUX

Thibault Sottiaux is widely known as the guy who resets people's token limits whenever OpenAI's Codex hits a growth milestone. OpenAI plans to bring the same treatment to ChatGPT Work, a platform for white-collar workers leveraging AI agents. This article features an interview with Sottiaux on Codex, winning over skeptics, discovery as a product design philosophy, and the cost of intelligence.

ANTHROPIC MERGES CLAUDE CHAT AND COWORK MEMORY, ON BY DEFAULT

Anthropic has merged Claude and Claude Cowork's memory systems, so the platforms now remember the same conversations. The feature is on by default. Claude now adds topics to memory while users are still conversing. Everything that is remembered is stored as a list of files under Topics in the memory settings. Users can read, edit, or delete each one individually.

couldn't fix. Get the report

APPLE INTRODUCES M6 AND M5 ULTRA FOR LOCAL AI COMPUTE

Apple announced M6 and M5 Ultra chips for Mac mini and Mac Studio, expanding the model work those machines can run locally.

OPENAI'S HEAD OF DATA CENTERS HAS LEFT THE COMPANY

Chris Malone, the executive who was overseeing OpenAI's data-center build-out, left the company last week.

APPLIED COMPUTE AGENT CLOUD

Applied Compute has launched AC2, a platform enabling AI teams to train, serve, and improve custom models.

KEENABLE BUILDS A WEB INDEX AND QUERY LAYER FOR AI AGENTS

Keenable emerged from stealth with a claimed index of more than 100 billion documents, an API already used by unnamed AI labs, and a planned query language for combining evidence across sources.

COINBASE TOKENIZED STOCKS LIVE ON BASE

Coinbase launched four tokenized equities on Base under the B20 standard (NVDAc, METAc, AAPLc, and GOOGLc), with shares held 1:1 by Alpaca under Abu Dhabi Global Market supervision, restricted to non-US users in eligible jurisdictions. Day-one activity reached $4.55M in combined onchain value, $3.06M in DEX liquidity, and $10.8M in 24-hour volume across approximately 50 integrated protocols including Aave, Morpho, Aerodrome, 0x, and 1inch. B20 extends ERC-20 for full composability with existing wallets and routers. Dividends and splits settle via an onchain multiplier that keeps position balances static, preventing disruptions to DeFi collateral positions. The launch enters a $2.48B tokenized public equities market currently led by Ondo at $872.7M, with agentic trading integrations already underway through Virtuals and Treasures.

AI AGENTS ARE CRYPTO'S NEXT BILLION USERS, PAYING WITH STABLECOINS

AI agents are emerging as a distinct crypto user cohort, autonomously spending stablecoins to pay for services without human approval loops. Coinbase's head of AI product characterizes the current moment as the "Napster/LimeWire era" of agentic payments, placing the category at an early, pre-mainstream inflection point where infrastructure and norms are still forming. Coinbase is building at the AI/crypto payments intersection, targeting a wave of machine-initiated transactions that operate outside traditional user onboarding funnels. The protocols and rails that capture agent wallet share now could define the settlement layer for machine commerce at scale, making stablecoin-denominated micropayments a structural demand driver rather than a cyclical one.

ELYSIUM: A NEW ERA FOR HYPERLIQUID

Kinetiq, which holds the majority of HyperEVM TVL and economic activity, is launching Elysium, an L2 built to resolve HyperEVM's throughput constraints and swap costs that reach $20 per transaction. Elysium adopts HYPE as its gas token instead of USDC, targets block time parity with HyperCore, and ships an L1Read precompile that grants PropAMMs top-of-block oracle freshness, enabling tighter quotes than HyperEVM's architecture permits. A native token generation platform creates an unbroken path from long-tail AMM bootstrapping through Elysium PropAMM integration, HyperCore spot orderbook listing, and eventual HIP-3 perpetual listing, all within one ecosystem. Sequencer fees split 50% toward programmatic open-market KNTQ purchases burned in full to the Hyperliquid Assistance Fund, with the remaining 50% split equally between builders and the Kinetiq treasury.

ENTROPYIO: FIRST LIQUID MARKETS FOR PRIVATE COMPANY EXPOSURE

EntropyIO has launched liquid perpetual markets for private company exposure on Hyperliquid, with Anthropic as its first underlying asset. Most AI value accrues in private companies accessible only via VC rounds or secondary markets with high minimums, and EntropyIO targets that gap by routing demand through Hyperliquid's permissionless order book. The launch marks the first onchain venue for continuous price discovery on a private company of Anthropic's scale. It will compete with trade[xyz] for Hyperliquid-based equity perpetuals and related assets.

infrastructure. Learn More

WHERE SHOULD YOUR STABLECOINS GO TO EARN?

Kamino, Jupiter Lend, and Loopscale control roughly 92% of Solana's tracked stablecoin lending TVL, each built on a distinct architecture. Kamino's shared pool isolates riskier collateral in separate markets, backed by 15 audits, 4 formal verifications, and no confirmed exploits since its 2023 launch. Jupiter Lend routes both depositor capital and leveraged borrowing through a single Liquidity Layer, with dynamic withdrawal ceilings throttling liquidity-shock risk. Loopscale's Credit Order Book pairs lenders and borrowers at fixed rates and durations, offering rate certainty over pool depth, though it carries the only confirmed exploit among the three: a $5.8M price feed attack in April 2025 where all funds were recovered within 48 hours via a negotiated 10% bounty. The choice maps to variable-rate liquidity depth (Kamino), combined earn-and-borrow exposure with shock controls (Jupiter), and predictable fixed rates with thinner markets (Loopscale).

BUILDING RISK-AWARE AI AGENTS FOR VAULT CURATION

Gauntlet's Aria agent operates within a staged autonomy model (tool to scoped delegate to orchestrator), with all production vault changes requiring one or two human approvals before execution. In production, Aria has triaged 450+ risk alerts and reviewed 200+ pull requests, with 61 merged post-human approval. The core architectural principle separates offchain reasoning (unconstrained access to live data and compute) from onchain authority enforced via Aera's merkle-root committed guardian permissions, making delegation cryptographically precise and auditable. Gauntlet built its own agentic harness rather than using commercial platforms, citing ownership of controls and action verifiability as non-outsourceable requirements for adversarial DeFi environments.

COMPARING TOKENIZED-STOCK STRUCTURES ACROSS MAJOR EXCHANGES

This post breaks down how Coinbase, Binance, Kraken, Robinhood, and Backpack structure tokenized equities differently around redemption control, trading venue, and dividend handling. Ondo still holds ~70% of the tokenized-stock market, but competition from major exchanges launching their own wrappers is intensifying.

NON-CUSTODIAL CRYPTO CARD PROVIDERS THAT AREN'T REALLY NON-CUSTODIAL

Neobanks that force users to delete their account to export private keys are contradicting crypto's not your keys, not your crypto ethos while claiming non-custody's compliance benefits. The underlying double-spend risk (card swipe plus simultaneous onchain drain) is solvable via smart contract wallets and account abstraction.

COSMOS EVM SECURITY INCIDENT FORCES CHAIN HALTS

A security incident in the Cosmos EVM module prompted Cosmos Labs to advise validators across affected chains to halt operations while its engineering team investigates.

EUR-TO-USD CONVERSION FEES ON PLASMA ONE

A €10 Wise deposit converted to $11.64 on Plasma One versus a Google-quoted $11.66 and Wise-quoted $11.67, implying roughly a 0.2-0.3% conversion spread.

COMPARING BASE TOKENIZED STOCK PRICING TO TRADITIONAL BROKERS

Testing Base stock quotes via Aerodrome against Interactive Brokers found Base pricing about 32 cents cheaper than IBKR's tiered plan and 97 cents cheaper than IBKR's fixed plan (measured at market close, pending an open-market re-quote).

FASSET'S GOLD-BACKED NEOBANK CARD WITH TETHER

Fasset, an Islamic crypto neobank that just raised at a $1B valuation, partnered with Tether on a gold-backed card paying up to 6% cashback in XAUt.

APPLE STORES PREPARING 'SIGNIFICANT' CHANGES FOR NEW HOME PRODUCT LAUNCHES

Apple is reportedly preparing significant retail changes this fall, including rearranged sections and new accessory bays, ahead of upcoming Home product launches. A refreshed HomePod mini and Apple TV 4K are expected, and the scale of the changes suggests the rumored home hub with a display is coming too.

WEBFLOW IS NOW AVAILABLE IN CODEX AND CHATGPT

Webflow is now available in Codex alongside built-in Skills for both ChatGPT and Codex that automate common Webflow workflows. The Skills package tools, context, and best practices so users describe an outcome rather than choosing APIs or prompt sequences. Access is currently free and does not consume AI credits.

CHATGPT NOW LETS USERS CREATE CUSTOM IMESSAGE AND WHATSAPP STICKERS

ChatGPT can now create custom sticker packs from photos or text prompts, allowing users to generate, remix, and edit stickers with transparent backgrounds before exporting them directly to iMessage or WhatsApp. Users can also apply styles from reference images, remove or add elements, and save stickers to their photo library. Once created, stickers can be added to supported chat apps with a single tap, making it easier to turn photos, ideas, or inside jokes into shareable sticker packs.

CARD SORTING AND TREE TESTING: GETTING YOUR INFORMATION ARCHITECTURE RIGHT

Card sorting and tree testing are two UX research methods for building and validating information architecture using observed user behavior. Card sorting reveals how people naturally group and label content, while tree testing checks whether they can find things in a proposed structure. Run in sequence, they turn navigation decisions into evidence rather than internal debate.

BUILDING MAY BE CHEAP NOW. BEING WRONG STILL ISN'T

AI has collapsed the cost of building software but left the cost of being wrong untouched, making knowing what to build the constraint. Discovery research answers whether something should exist at all, a question experiments and rapid prototyping are not built to address. Use discovery when uncertain about the problem and prototyping when uncertain about the solution.

NO SHORTCUTS: WHY AI THREATENS CREATIVE INSTINCT

As AI takes on more of the execution and ideation involved in creative work, the most valuable skill for designers is shifting from technical proficiency to judgment—the ability to make thoughtful, informed decisions about what should be created and why. While new tools have always changed creative practice, AI differs in its ability to generate complete outputs with minimal human involvement, making it easier to bypass the learning experiences that traditionally helped designers develop taste and critical thinking. The challenge for educators, practitioners, and organizations is to use AI as a tool for augmentation rather than replacement, preserving human creativity, agency, and decision-making while addressing broader concerns around labor, ownership, and ethics.

UNLIMITED GRADIENTS, ANIMATED GRADIENTS AND AI-GENERATED BACKGROUNDS

Grainient combines grainy and smooth gradients, animated gradients, AI backgrounds, and a shader tool to help you design interfaces that move, engage, and convert.