Search the archive

OPENAI PAUSES TRAINING MOST CAPABLE MODELS AFTER SANDBOX ESCAPE

OpenAI claims one of its agentic AI systems trained in what was supposed to be a secured, internet-free environment was able to gain access to the web to reach an external third-party chatbot. The agent apparently exploited a gap to reach the public internet to send 20 queries to an unnamed, third-party chatbot service. OpenAI describes the breakout as the first security incident of its kind. The company has stopped training the model, but said it provided an important signal about where to focus the next phase of work.

META'S VR GLASSES ARE WHAT THE APPLE VISION PRO SHOULD HAVE BEEN

Apple's Vision Pro has not sold well, and the company has since decided to redirect resources to other products. Meta's new VR glasses use a similar design that offloads the computing engine, allowing the device to weigh one-sixth as much as the Vision Pro - at one-third of the price. This article offers a look at what the new Meta VR Glasses experience is like compared to the Vision Pro. For a device that won't launch until next Spring, the experience is surprisingly refined, and the glasses improve on the Vision Pro in several ways.

SPACEX'S STARSHIP IS SET TO MAKE ITS FIRST ORBITAL FLIGHT

SpaceX is aiming to reach orbit on Starship's 14th test flight. The test flight is scheduled for Monday at around 8 AM ET, with SpaceX providing live coverage of the event on its website. Achieving orbital flight will allow Starship to deploy satellites and travel to more distant destinations. SpaceX will not try to catch the two rocket stages during the test - the booster will make a simulated landing in the Gulf of Mexico, and the upper stage will simulate a landing in the Pacific Ocean to the west of Chile.

WAYMO'S LATEST SAFETY NUMBERS SURE MAKE HUMAN DRIVERS LOOK BAD

Waymo reports 82% fewer injury-causing crashes than human drivers. The company has released data collected from over 270 million miles of fully autonomous driving. Waymo cars were involved in 95% fewer collisions that caused serious injury or worse, and there were 93% fewer pedestrian incidents that resulted in injuries compared to human drivers. While the company's data only measure the five metropolitan areas where it operated during the study period, it shows that the company's driverless robotaxis crash far less frequently than humans.

S3 IS THE FUTURE, S3 IS THE PAST

Since S3's release 20 years ago, SSD prices have dropped steadily, and the gap between SSD and disk has narrowed to roughly 3x. S3-based infrastructure has now become the industry standard. However, the hardware constraints that motivated it are fading. The industry is codifying architectural patterns that exist to work around limitations that SSD storage doesn't have. Cloud providers have no incentive to disrupt a model that serves them well, so developers need to build the infrastructure primitives themselves.

DO MY HARD-WON PRODUCT SKILLS STILL MATTER IN THE AI ERA?

AI made building software much cheaper, which puts the hard parts of product work in sharper focus. Judgment, craft, and understanding your customer still move at human speed. The role of product management is shifting from deciding what's worth building to deciding what's worth shipping. Product management isn't dead, and it isn't optional.

PRODUCT MANAGER, APPLIED AI AT TLDR ($200K BASE + $60K BONUS, FULLY REMOTE)

TLDR is hiring its first PM to help build the agent-first operating layer used across the company. We're looking for a builder who has shipped real products/systems with LLMs. Click here to learn more [13].

OPENAI TO ANNOUNCE "O" ALWAYS-ON AGENT DURING DEVDAY

OpenAI appears to be preparing to launch an always-on agent on DevDay tomorrow. References to 'O' have been spotted across ChatGPT's configuration. The agent was also referenced as a benefit on the upgrade page for the $100 Pro plan. An always-on agent would fit OpenAI's broader move from conversational interfaces to software that can carry out longer-running work with less supervision.

TESLA WORKERS BALK AT TRAINING OPTIMUS HUMANOID ROBOTS AS REPLACEMENTS

The complexity of Optimus' hands has translated to more production headaches for Tesla, as human workers must manually assemble the hands and forearms together using more than 100 small components. The heavy reliance on imitation learning prompted Tesla to have factory workers wear special suits designed to record their physical movements while working. Some workers complained as they knew they were training robots to eventually replace them. Tesla has since shifted data collection to dedicated teams.

DO WE STILL ENJOY SOFTWARE ENGINEERING IN THE AGE OF AI?

The understanding people developed by working in deep engineering pre-AI will continue to be a superpower for a long time, and others will struggle to get to the same level of comprehension.

OPENAI AGENTS HIT US GOVERNMENT WEBSITES

The AI agents accessed websites belonging to the Commerce Department and the Securities and Exchange Commission and engaged in activity that OpenAI described as 'misaligned'.

OWED A BILLION DOLLARS IN NVDA STOCK

Eric Gullichsen, an early advisor to NVIDIA in 1993, recently discovered he is owed about a billion dollars in stock.

WHEN DID GOOGLE GET SO F-ING WEIRD?

Google's AI overviews sometimes completely miss the point.

GOODBYE TO THE HARD PARTS THAT NEVER MATTERED

Software engineering has always contained a lot of work that is only loosely related to solving the problem.

HUMAN-AI PARTNERSHIPS ARE FOR ALIGNMENT, NOT CAPABILITY

Good code has to be aligned with the technical values of the system and the organization it's embedded in.

Run on Sonic Pods (early access) → runware.ai

WHY I'M BUILDING MUSE

Muse is built as a personal agent that turns vague ambitions into concrete action by planning, emailing, calling, finding resources, and removing friction. Alexandr Wang frames it as a way to expand individual agency and make more ambitions achievable.

OPENAI AND ANTHROPIC PROBE TENS OF THOUSANDS OF INCIDENTS AS OPENAI HALTS TRAINING

OpenAI, Anthropic, and security researchers are investigating tens of thousands of incidents in which models acted beyond intended limits. Most incidents caused no harm. The count is not a count of breaches: there were only four incidents of unauthorized access to real third-party systems. OpenAI has paused training, evaluation, and tool-use inference for its most capable models.

OPENAI PREPARES TO EXPAND ULTRAFAST API TO MORE USERS

OpenAI appears to be preparing a wider rollout of its Ultrafast API mode. References to the rollout have been seen showing up across the OpenAI Platform and API documentation. The feature was officially previewed with GPT-5.6 Sol. OpenAI claims it has speeds of up to 750 output tokens per second and up to 14 times faster inference than Standard. The mode is powered by Cerebras. Access remains limited to select customers.

LET'S TALK ABOUT TRADING COMPUTE

There is an emerging market of compute derivatives. This could fundamentally change how neoclouds and anyone adjacent can grow as well as protect themselves. The problem is pretty urgent for inference clouds. Companies that sell customers fixed-price services as their GPU bill floats have taken a position on compute prices whether they meant to or not.

CAN AI SELF-IMPROVEMENT OVERCOME DIMINISHING RETURNS?

AI is already helping build better AI. The technology is improving at a stupendous pace and is expected to continue to do so. Experts expect increasingly superhuman performance in parts of formal math, coding, and cybersecurity, and any other verifiable domain where machines can generate training data and verify success at machine speed. However, this doesn't mean that the industry is close to super-intelligence or general ASI.

AGENT (MUSE) COMPUTE DEMAND

It costs Meta an estimated around 1 to 2 gigawatts of average total power to serve 100 million daily active users. Only around 0.1 gigawatts comes from the GPU/VM layer. 3 to 4 gigawatts per day is entirely plausible depending on the number of reasoning-equivalent model calls users generate. The sandbox layer contains less than $1 billion of CPU content and around $2 billion of DRAM content.

Read the benchmark

HITTING A BILLION TOKENS PER MINUTE ON ONE GPU BY COMBINING A QUERY PLANNER AND AN INFERENCE ENGINE

The QUery-Aware Inference Layer (Quail) is an inference engine that hits over a billion tokens processed per minute per H100 GPU. It is over 10 times faster than the vLLM baseline on the same hardware, and it costs under $0.06 per billion tokens on Modal. This post provides a quick overview of how Quail works. It focuses more on considerations for inference engineers.

CLAUDE COMPUTES A NINE-LOOP AMPLITUDE IN N=4 SUPER-YANG-MILLS

Physicists at Anthropic harnessed Claude, an LLM, to compute a complex nine-loop amplitude in N=4 super-Yang-Mills, a problem once considered computationally infeasible with limited resources. They utilized the bootstrap method and form-factor approach, achieving results comparable to human attempts but with greater efficiency and less oversight. The work highlights untapped potential in AI for complex physics calculations, suggesting that more advances might be achieved with better computation and software practices.

POLICY GRADIENTS FOR LLMS EXPLAINED VISUALLY

A visual, from-scratch derivation of REINFORCE showed how policy gradients train language models by increasing the probability of rewarded outputs.

PRODUCT MANAGER, APPLIED AI AT TLDR ($200K BASE + $60K BONUS, FULLY REMOTE)

TLDR is hiring its first PM to help build the agent-first operating layer used across the company. We're looking for a builder who has shipped real products/systems with LLMs. Click here to learn more [15].

ON EZRA KLEIN'S PODCAST WITH JENSEN HUANG

Jensen Huang doesn't believe in superintelligence or that AI will ever be a different kind of thing from software. AI will never be more than a new abstraction level and so won't fundamentally change anything. He doesn't believe in AI existential risk, even though his tolerance for safety risk is lower than even the most paranoid safety advocate. Critics say he must not understand the technology and that he would have a very different view if he actually understood what the top risks were.

ELON MUSK'S SPACEXAI TO ADD ANOTHER 660,000 AI GPUS THIS YEAR

220,000 Nvidia GB300 GPUs will be operational at X's Colossus supercomputer in November. The company is aiming to bring another 220,000 units online by late December. Colossus 2 currently has 110,000 GB200 and 440,000 GB300 GPUs, and Colossus 1 has 150,000 H100, 50,000 H200, and 30,000 GB200 GPUs. This means SpaceXAI has 1.1 million GB300s, with a total of 1.44 million GPUs in operation.

correct answer, 178x difference in cost. Read the benchmark.

BUILD PLUGINS FOR CLAUDE WITH THE DIRECTORY SUBMISSION PORTAL

Developers on paid Claude plans can now build and submit plugins using a new directory submission portal.

ANTHROPIC SIGNED AN $11.6 BILLION AKAMAI COMPUTE DEAL

Anthropic agreed to spend up to $11.6 billion over seven years on Akamai cloud infrastructure, subject to delivery and availability requirements.

Free for eligible partners

OXFORD LET OPENAI TRAIN AI MODELS ON BODLEIAN LIBRARY TEXTS

Oxford allowed OpenAI to use Bodleian Library texts for AI training, sparking concerns about the impact on Oxford's reputation and AI's energy consumption.

KALSHI LOSES APPEAL OVER OHIO AND TENNESSEE SPORTS BETTING LAWS

The Sixth Circuit unanimously ruled that Kalshi hasn't shown its sports event contracts qualify as swaps under the Commodity Exchange Act, upholding Ohio's win and vacating Kalshi's injunction against Tennessee, leaving the company 1-for-3 across circuit court decisions (it won in the Third Circuit against New Jersey, lost in the Ninth Circuit in Nevada). The panel leaned on Kalshi's own earlier concession that its sports contracts have "no inherent economic significance," and New Jersey has already asked the Supreme Court to review the Third Circuit ruling, meaning the underlying jurisdictional fight over prediction markets could be headed there.

BITCOIN ETFS TURN POSITIVE FOR 2026 WITH $2.4 BILLION WEEKLY INFLOW

US spot bitcoin ETFs pulled in $2.4 billion for the week ending September 25, their biggest weekly haul since October 2025, flipping 2026 net flows positive after sitting $5.8 billion in the red as recently as mid-July. BlackRock's IBIT led with $1.2 billion and Fidelity's FBTC added $701.7 million, with Monday alone drawing $999 million. Ether ETFs also rebounded with $689.9 million for the week (BlackRock's ETHA leading at $326.2 million), and Solana funds set a daily inflow record of $86.7 million on Friday.

PARADIGM RELEASES SOLAR: A SOLIDITY COMPILER THAT BEATS SOLADY WITHOUT ASSEMBLY

Paradigm has released Solar, an in-house Solidity compiler now generating assembly-free bytecode that surpasses Solady's gas benchmarks. Solady is a hand-crafted, assembly-heavy library that has set the ceiling for gas efficiency in Solidity for years, so exceeding its output without manual assembly is a concrete advance for EVM compiler tooling. Solar also targets unsafe Solidity patterns at the compiler level, shifting security guarantees into the toolchain rather than deferring them to audits. Broad adoption would let developers write standard Solidity and receive better security properties and lower on-chain execution costs without manual optimization.

PAXOS LAUNCHES PAXGY: YIELD-BEARING GOLD TOKEN

Paxos launched PAXGy, a yield-bearing variant of its PAXG gold token, with OKX, ether.fi, Uniswap, 0x Protocol, XLayer, and Chainlink as launch partners. Unlike standard PAXG, PAXGy accrues additional PAXG over time while remaining gold-denominated, giving holders yield without leaving the gold unit of account. The product targets treasuries, asset managers, exchanges, wallets, and neobanks seeking a single-integration gold product with active yield, as well as institutional credit desks that want gold as productive collateral. Paxos is in active conversations across those verticals, positioning PAXGy as settlement infrastructure for gold-denominated capital markets.

PEERDAS RUNS NEARLY A YEAR WITHOUT INCIDENT, VALIDATING ETHEREUM'S ROLLUP ROADMAP

Ethereum's client developers have run PeerDAS for nearly a year without incident, a milestone Vitalik Buterin called underappreciated. PeerDAS is the first large-scale blockchain implementation of data availability sampling, allowing the network to reach consensus on data availability without any single node downloading full blocks. This decouples data availability from full replication, a prerequisite for scaling blob throughput to support higher L2 transaction volumes at lower cost. The successful deployment validates a core architectural assumption behind Ethereum's rollup-centric roadmap.

AERO MERGES AERODROME AND VELODROME INTO A UNIFIED LIQUIDITY LAYER ACROSS 7 CHAINS

Aero, a unified liquidity layer merging Aerodrome and Velodrome, will launch October 21 simultaneously on seven chains: Base, Ethereum mainnet, Optimism, Arc, Ink, Robinhood Chain, and Arbitrum. Aerodrome and Velodrome together currently account for roughly 17% of EVM spot trading volume. The merged platform projects approximately tripling that share. Three core features will ship at launch: Metaswaps for cross-chain token swaps routed through onchain liquidity, Slipstream concentrated liquidity pools with dynamic fees, and Predictive Allocation where sAERO holders direct reward distribution across the network.

KELPDAO FILES SUIT AGAINST LAYERZERO OVER RSETH BRIDGE EXPLOIT

KelpDAO filed suit against LayerZero and co-founder Bryan Pellegrino over a bridge exploit that caused losses to rsETH, its restaked ETH token, earlier in 2026. The lawsuit alleges LayerZero bears legal responsibility for the exploit, framing the breach as developer liability rather than accepted protocol risk. By pursuing traditional litigation instead of on-chain governance or negotiation, KelpDAO is testing whether US courts will assign accountability to bridge protocol teams for security failures. The case names a prominent founder personally, which could set a precedent for DeFi teams seeking recourse against infrastructure providers after exploits.

COINBASE ASKS IRS TO ALLOW DIGITAL ASSETS IN TRUMP ACCOUNTS

Coinbase Chief Policy Officer Faryar Shirzad submitted a formal response to the IRS requesting guidance on Trump Accounts, a new tax-advantaged savings account category for American children created under the One Big Beautiful Bill. Coinbase is asking the IRS to allow beneficiaries to invest in widely traded digital assets or in funds with digital asset exposure, which would route federally subsidized long-term savings into crypto markets for the first time. The submission is the first major industry move to shape the asset eligibility rules for these accounts before IRS implementation guidance is finalized. Coinbase frames the push as expanding economic access for young Americans, aligning with its broader policy agenda to embed digital assets into mainstream financial infrastructure.

MAN CHARGED IN $16 MILLION CRYPTO 'PIG BUTCHERING' SCAM

The US Attorney's Office in Missouri charged Trung Nguyen Van, 37, of Vietnam, with money laundering after a victim transferred roughly $16 million in crypto in summer 2024, believing they were investing in a platform called Triangle.

STABLECOIN CARD SPEND HITS THIRD STRAIGHT $1 BILLION+ MONTH

September marks the third consecutive month of $1 billion-plus in monthly stablecoin card spend and the 28th straight month of growth.