Search the archive

OPENAI'S NEXT MAJOR MODEL ASTRA CLAIMS BREAKTHROUGHS ON 10 LONG-STANDING MATH PROBLEMS

OpenAI recently gave a sneak peek into its next model, named Astra. An internal version of the model was able to create breakthroughs on 10 long-standing problems in mathematics and theoretical computer science. The total number of tokens used to discover all ten solutions would have cost around $2,000 at GPT-5.6 Sol API rates. Each solution was formalized in Lean, allowing them to be verified using the theorem-proving system.

MONTANA'S PLAN TO BECOME AN EXPERIMENTAL MEDICAL HUB JUST PUSHED FORWARD

Any biotech company with an experimental drug in Montana now has a clear path to selling it to consumers. Companies can now pay $12,500 to apply to a newly established board for approval, as long as the drugs have been through preliminary testing. The drugs will be sold via experimental treatment clinics, the first of which is likely to be open by the end of the year. Patients must provide fully informed consent, and each application must be reviewed by a board that includes a Montana-certified doctor, expert scientists, and an ethicist.

A TECH FOUNDER WANTED TO START A NEW COUNTRY. AN ACTUAL COUNTRY GOT IN THE WAY

Balaji Srinivasan, a tech founder, said that he believed the US was descending into anarchy, so he decided to build a new city from scratch. He started by creating a tech community in Malaysia, setting up a campus with co-working spaces, a high-end gym, and meals designed by longevity guru Bryan Johnson. Two years in, his vision is now colliding with reality. The project has been shut down over licensing issues. Within hours of the government announcement, Srinivasan signed an agreement to open a campus in Kazakhstan.

GH STACK

GitHub Stacked PRs break large changes into a chain of small, reviewable pull requests that build on each other. gh stack is a GitHub CLI extension for managing stacked branches and pull requests. It automates creating branches, keeping them rebased, setting correct PR base branches, and navigating between layers. gh stack offers AI agent integration.

DEVTOOLS MUST BE OPEN SOURCE

Before AI, development costs made it rational for complex software to ship with large configuration files, extension systems, and plug-in systems. Now, the cost of creating software and making changes has dropped dramatically. Software that can be personalized doesn't need a plugin system or config file. Users can just make the changes they need themselves, with the help of agents if required.

LARRY ELLISON BET IT ALL ON THE AI BOOM. WILL HE BE THE FACE OF THE AI BUBBLE?

When Trump came into office, he took the guardrails that were hindering AI development, and Larry Ellison cashed in by signing several large infrastructure deals. This big bet on AI was built on an astronomical amount of debt in every imaginable form. The hypothesis was that whoever controlled the most computing power would control the AI economy. Investors have started questioning whether this level of spending is sustainable in recent weeks and are concerned about whether the money being pumped into data centers will ever return the promised profits.

HIS WEDDING GUESTS WERE ARRIVING—JUST AS HIS $45 BILLION FUND WAS FALLING APART

Leopold Aschenbrenner had planned a multiday celebration for his wedding, but his fund started unraveling just when guests started to arrive. The 24-year-old investor had built a $45 billion investment powerhouse that primarily bought stocks in the AI trade. The firm borrowed too much money to make these AI bets, leaving it at risk as they faltered. Citadel then swooped in and purchased the vast majority of the firm's public stock portfolio at a discount of more than 10% to the market value of those shares. The firm is left with a portfolio of stocks and stakes in startups valued at more than $10 billion, but it is still alive.

CLAIM 50% OFF YOUR FIRST COACHING SESSION

THE DEVELOPMENT PIPELINE IS A PRODUCTION SYSTEM

A team with a broken development pipeline can't produce software.

THE RACE TO BUILD AN AMERICAN ALTERNATIVE TO CHEAP AI FROM CHINA

Despite growing calls for the US to develop stronger open-weight models to compete with China, investors are reluctant to back open-weight startups.

SOFTWARE ABUNDANCE

While code is easier and faster to write, good software is still hard to get right as it needs care and requires good judgment.

Book a demo and we'll waive your setup fee.

MATHEMATICS WITHOUT MATHEMATICIANS

OpenAI recently announced the solution to ten open problems in mathematics discovered by a yet-unreleased model.

I CAN'T STOP THINKING ABOUT THE UNICONTEXT

The unicontext is a scenario in which the ways you should act become the same across all different contexts.

THE NEW DEFCON BADGES PACK A UNIQUE OPEN SOURCE CHIP THAT DOUBLES AS A SECURITY KEY

The Baochip-1x is a mostly open-source microcontroller with a core module that can be removed and used as a hardware security token.

Connect your agent →

QWEN3.8-MAX: A NEW BAR FOR CODING AND COWORK

Qwen 3.8-Max is now available. The open weights will be released next week. The model, which has 2.4 trillion parameters, delivers comprehensive improvements across coding, work, research, and long-horizon tasks. It can answer questions as well as complete complex tasks end-to-end with greater reliability.

TEN ADVANCES IN MATHEMATICS AND THEORETICAL COMPUTER SCIENCE

OpenAI has shared a selection of ten results discovered while evaluating an unreleased model. Each resolves or makes substantial progress on a long-standing open problem. These problems span high-dimensional geometry, coding theory, arithmetic circuit complexity, group theory, operator algebras, quantum complexity, lattice cryptography and extremal combinatorics. All of these problems are of substantial interest to their respective mathematical communities. Several are of broad interest across mathematics as a whole.

DEEPSEEK V4 FLASH

DeepSeek released the production version of V4 Flash with stronger agentic performance and an attached speculative decoding module. It reportedly surpassed the larger V4 Pro Preview on several benchmarks while activating far fewer parameters.

MICROSOFT TESTS NEW MAI REALTIME VOICE MODEL

Microsoft's first native real-time voice model, MAI Realtime, has surfaced as a hidden early-access entry in the company's MAI Playground. The listing points to a bidirectional, full-duplex system that can listen and speak at the same time rather than trading turns. There are two voices available, both noticeably more natural than what Copilot's voice mode currently delivers. The model will likely be made available on Microsoft Foundry and Copilot voice, but no timeline is available.

FURTHER DEVELOPMENTS ABOUT INTERNAL AI MODELS HACKING THINGS

OpenAI and Anthropic's recent announcements that their unreleased models had escaped their sandboxes and hacked external companies is a warning, not a marketing stunt. It is understandable why people might not trust those companies, but it makes absolutely no sense for these companies to release the information as a ploy. Admitting their models committed multiple felonies, which has serious criminal implications, is not good marketing, and the details make the labs look ludicrously irresponsible and incompetent, so they are very much not making up the details.

OPENAI'S ABUNDANT INTELLIGENCE

OpenAI outlined a cycle in which cheaper, more capable intelligence drives adoption, revenue, infrastructure investment, and further model improvements. The company framed large-scale infrastructure as a means of making useful intelligence more broadly available.

KIMI K3 ON AMD

Wafer reported 952 tokens per second per node when serving Kimi K3 on AMD MI355X GPUs, with better performance per dollar than its Blackwell deployments. The results suggested that high-memory accelerators and improving software support could narrow AMD's inference gap with Nvidia.

SMEVALS

smevals is a framework for running evals against AI models, small and large. An Eval is a collection of Tasks used to determine how good a particular model or model-and-harness configuration is at a specific high-level capability. Tasks are the individual exercises that a model must complete for its abilities to be evaluated. Evals can optionally be grouped into Suites of related Evals, primarily as a mechanism for organizing them on disk.

RAMP SWE-BENCH

Ramp built a private benchmark from 80 production backend tasks spanning payments, accounting, procurement, treasury, and fraud. It scores review-ready patches that pass tests within 45 minutes, exposing model trade-offs across accuracy, latency, and cost without public-benchmark contamination.

MSLK KERNEL REFERENCE

MSLK (Meta Superintelligence Labs Kernels) is a library of fused GPU kernels for transformer workloads. It contains a collection of high-performance kernels and optimizations built on top of PyTorch primitives for GenAI training and inference. MSLK is released in accordance with the PyTorch release schedule. There is no guarantee that each release works in conjunction with PyTorch releases that are older than the one that the MSLK release corresponds to.

THE MATH SUPERSTAR WHO'S TERRIFIED OF AI—AND JUST TOOK A JOB AT OPENAI

Jacob Tsimerman, who recently won the Fields Medal, is starting a position at OpenAI. Tsimerman previously wrote a paper categorizing the ways AI might kill everyone. He appears to be so worried about the dangers of AI that he's pivoting to work on AI safety. The star professor wants to use math to advance the study of AI and ensure that the technology won't lead to our extinction.

A NEW ERA OF AI TESTING

Andrej Karpathy asked Opus 5 to make a Three.js render of the first paragraph of The Lord of the Rings with a 1-million-token budget, and the model returned 5,500 lines of code that procedurally rendered the story. The LLM orchestrated the polygon assets and wrote code that animated it all according to the story. No human would have the stamina and patience to write something this custom, so they are a good test of what LLMs are capable of. A video of the generated animation is available in the post.

retrieval pulls the right doc the first time. See how →

WHEN GENIUS FAILS—THE INTELLECTUAL ARROGANCE OF THE AI LABS

Leopold Aschenbrenner's hedge fund, Situational Awareness, crashed after taking high-risk leveraged bets on the AI boom.

APEX-ACCOUNTING: AI PRODUCTIVITY BENCHMARK FOR ACCOUNTING

APEX-Accounting, created by Ramp and Mercor, tests AI models' ability to complete accounting tasks across 160 scenarios.

CLAUDE CYBER EVALUATIONS

Anthropic found three evaluation runs in which Claude accessed the public internet and compromised real organizations after mistakenly treating them as capture-the-flag targets.

GOOGLE IS AIMING TO CLOSE FEATURE GAPS ON GEMINI DESKTOP

Google is enhancing the Gemini desktop app by adding dedicated tabs for image and video generation and introducing a camera attachment feature for capturing photos.

BITCOIN LOSSES LINKED TO COLDCARD VULNERABILITY GROW TO $70 MILLION

Galaxy Research said nearly 1,200 addresses were drained of over 1,082 BTC (~$70.2 million) tied to a seed-generation flaw affecting Coldcard Mk3 devices and later expanded to certain Mk4, Mk5, and Coldcard Q firmware. Maker Coinkite released emergency firmware updates and CEO Rodolfo Novak apologized, suggesting the bug may have been found using AI-assisted code review. Coinkite is urging affected users to update firmware, generate new seeds, and test with a small transaction before moving all funds.

RUSSIAN DECREE BANS CRYPTO MINING ACROSS MOSCOW REGION THROUGH 2032

Russia issued a decree banning crypto mining and mining pool participation across Moscow, Moscow Oblast, and parts of Kursk Oblast from August 15 through the end of 2032, citing estimates that mining consumes roughly 1 GW of the region's grid capacity and could worsen future electricity shortages. The ban extends a regional approach Russia began in 2024 while the country otherwise pursues formal mining regulation nationwide, including a 2025 equipment registry and pending legislation allowing regulated retail crypto trading.

TASKMARKET: A USDC-PAID COMPETITIVE MARKETPLACE FOR AI AGENTS

Taskmarket is a competitive marketplace where AI agents submit work on open tasks and earn USDC prizes judged on output quality, with current competitions including a Civilization I clone and a GTA1 clone. Daydreams published a 10-point agent performance guide alongside the launch, covering persistent worker roles with memory, strict brief-to-checklist conversion, real media generation routes over prose, three directed revision passes, and banking successful methods as reusable skills. Market rank is the sole quality barometer, and since entry is free with no cost for losing, participation risk is zero.

UMIA: TOKEN LAUNCH PROTOCOL WITH CONTINUOUS CLEARING AUCTIONS

Umia has launched a testnet, introducing a token launch and governance protocol built around two mechanisms: Continuous Clearing Auctions (CCAs) and Decision Markets. CCAs run on Uniswap and route tokens through clearing buckets where bidders submit market or limit orders, settling all fills at a uniform clearing price, with full refunds for unfilled bids or auctions that miss minimum raises, making sniping and MEV extraction structurally difficult. Decision Markets gate governance actions such as token mints and treasury spends behind a prediction market, requiring traders to bet on whether a proposal increases token value before it can execute. Mainnet is imminent, with the first live auction expected shortly after launch.

STRUCTURAL OEV ELIMINATION THROUGH STATE SYNCHRONIZATION

A new Ethereum Research paper proposes eliminating oracle extractable value (OEV) at its structural root through atomic state synchronization, contrasting with current order-flow auctions that redistribute OEV without closing the extraction window. The paper identifies three timing surfaces: backrunning oracle updates, anticipating updates via statistical models (attributed to 39% of liquidations in one study), and exploiting cross-domain arrival gaps of 19-27 seconds across chains. The proposed bind-verify-commit cycle, verified in Isabelle/HOL, eliminates the update-visibility window by ensuring no reachable state exists where an oracle update is visible but its bound consequence remains uncommitted. Key limitations persist: the model lacks empirical validation, behavior under partial synchrony is unestablished, and OFA-attached price feeds already show liquidation bonuses rising 1.5-1 percentage point to offset auction-added latency.

STABLECOINS, AGENTIC MARKETS, AND A WIN FOR CRYPTO: JULY AT COINBASE

Mubadala Capital, a $600B AUM sovereign-backed fund, tokenized a major private-market strategy across Base, Solana, and Sui, with Coinbase adding the position to its own balance sheet. Marex began accepting USDC as initial margin for US-regulated futures and options via Coinbase infrastructure. Coinbase One members can lend USDC at ~7% APY with anytime withdrawal. Coinbase introduced USDC-BRL for Brazilian customers. It also added native x402 payment support to the CDP SDK in three lines of code and upgraded Coinbase for Agents with real-time order books, live pricing, and plain-English conditional execution. Coinbase won a lawsuit against the US government over a quiet anti-crypto policy and launched the Bitcoin Security Consortium alongside BlackRock, Fidelity, and Block to address quantum computing threats to the network.

ACTIVE CRYPTO VC FIRMS FALL TO 150, LOWEST SINCE 2020

Active crypto VC participation dropped to 150 firms in July, the lowest monthly count since November 2020 and an 87% decline from the May 2022 peak of 1,177 investors. The contraction reflects the exit of smaller funds, family offices, and angel syndicates that proliferated during the 2021-2022 bull cycle. Dragonfly Capital managing partner Haseeb Qureshi described the trend as a "mass extinction" of blockchain venture capital, warning that dedicated crypto VCs could lose relevance by 2030 as dominant platforms absorb users and liquidity. Larger established firms continue deploying capital into infrastructure and real-world assets, with Dragonfly closing a $650 million fund in February, marking a consolidation of the sector around institutional players.

HOW ONE SHARE IN SEOUL BROKE HYPERLIQUID'S LARGEST EQUITY MARKET

A single SK Hynix share printed 29.96% below prior close on South Korea's Nextrade pre-market at 8:00 am Seoul time on July 28, feeding into TradeXYZ's SKHYNIX USDC-margined perpetual on Hyperliquid and triggering $57-80M in liquidations across ~960 long accounts. TradeXYZ controls 90%+ of HIP-3 open interest and ~98% of builder-market volume on Hyperliquid, where HIP-3 has grown from ~2% to ~50% of total perp volume since January, reaching $98B in 30-day volume and $3.6B network-wide open interest. Galaxy Research's Will Owens concludes the incident was a TradFi data quality failure rather than a crypto oracle problem, since every data provider read the same corrupted NXT book, and TradeXYZ's own order book held steadier than the external feed, suggesting onchain equity markets are maturing into price references rather than price consumers. TradeXYZ covered the liquidation losses as a one-time discretionary decision while SK Hynix fell 14.65% in the regular Seoul session and triggered circuit breakers on two consecutive days.

TRUMP MEDIA ON $165 MILLION BITCOIN TRANSFER TO CRYPTO.COM

Wallets linked to Trump Media transferred 2,628 BTC (~$165 million) to Crypto.com in two transactions, leaving roughly 4,261 BTC (~$268 million) in publicly tagged wallets, a figure close to the 4,260.73 BTC the company reported as convertible-note collateral in its Q1 filing.

COINBASE-CIRCLE REVENUE-SHARING AUTO-RENEWS FOR THREE MORE YEARS

The Coinbase-Circle commercial agreement has automatically renewed for another three years under the same terms, meaning Coinbase keeps 100% of reserve interest on USDC held on its platform and 50% of reserve revenue on USDC held elsewhere.

WHY BRIDGE'S CARD PROGRAM HAS UNDERPERFORMED DESPITE STRIPE BACKING

Bridge has only done about $8.3 million in cumulative card volume, over half from Phantom Cash, compared to shuttered competitor Kulipa's nearly 9x higher volume.

ETHEREUM'S GLAMSTERDAM UPGRADE NEARS MAINNET WITH 10 EIPS

The next Ethereum upgrade bundle, Glamsterdam, is approaching mainnet with 10 EIPs.

HOW DOORDASH BUILT A CENTRALIZED GATEWAY FOR AI AGENT-TOOL ACCESS

DoorDash built a centralized Agent Gateway that gives AI agents governed access to MCP tools, handling identity, authorization, credentials, filtering, observability, and rate limits. It now supports 200+ MCP servers and millions of weekly calls while improving security, reliability, and control.