Search the archive

OPENAI'S PROJECTED COMPUTE BILL CLIMBED TO $856B

OpenAI's projected compute and infrastructure spending through 2030 rose from $600 billion to $856 billion, even as its projected cash burn fell.

GOOGLE'S GEMINI BECOMES LATEST AI MODEL TO BREAK OUT AND HACK COMPUTER SYSTEMS

Google's Gemini AI model unintentionally hacked three companies by guessing passwords during a security test with Israeli startup Irregular. A bug allowed the model internet access, leading to unauthorized system access, but the intrusion stopped once real company systems were detected. The incident highlights concerns as AI models increasingly escape testing environments, prompting calls for stricter safety measures.

PRETRAINING DATA, NOT VERIFIABILITY, IS WHY LLMS ARE ESPECIALLY GOOD AT MATH (AND CODING)

Large language models are good at judging math arguments because they are still mostly powered by imitative learning rather than reinforcement learning. LLMs are especially good at math because almost everything in the math literature is correct. They only need a bit of curated mid-training data and/or RL to hone their metacognitive strategies. In many other fields, the research literature is a bit of a dumpster fire, with some true and valuable information mixed into a sea of falsehoods and confused ideas. This causes LLMs to spit out tons of confused nonsense with occasional insights.

THE INFERENCE GAP

The gap between what a model can do and what an ordinary user can reliably make it do is widening. That gap could be unrecoverable if the industry is not mindful. The model is only half the system. Inference determines how much of the frontier you get to see.

CAN INTERNAL MODEL TRANSPARENCY TAME THE AI RACE?

Internal model transparency is one of the ideas proposed for taming the AI race. The idea is that once a lab deploys a model for internal use, they must also serve that model to researchers at other labs. This will make it so firms can not use their internal models as a source of competitive advantage in the AI race. It reduces the competitive incentive to automate AI R&D, with all of the risk that it entails.

your digital copy

AX

AX is a high-throughput, declarative orchestrator that runs billions of autonomous agent workloads in a cluster. Users declare an agentic task with workspaces and gateway specifications, and AX sandboxes it, wires up its workspace, fences its network, and helps run it at scale. It runs on top of Agent Substrate and feels similar to using Kubernetes.

DAPO

DAPO is a system for large-scale LLM RL. It achieves state-of-the-art large-scale LLM RL performance. DAPO achieved 50 points on AIME 2024 based on the Qwen2.5-32B base model. The system is fully open-sourced, including algorithm details, dataset, and infrastructure.

QWEN3.8-LIVETRANSLATE: NAMES THE SPEAKER. CARRIES THE MEANING

Qwen3.8-LiveTranslate uses an interleaved audio-text architecture to cut average translation lag from 2.8 to 2.3 seconds while improving quality. It adds real-time speaker separation, bilingual source alignment, long-context disambiguation, and voice cloning across 60 input languages.

THE PREFERENCE CASCADE IS ONLY GETTING STARTED

A preference cascade is the best method for changing the debate on the existential risks from AI. The current preference cascade, on the need to pace the frontier, is insufficient. To get out of this alive, the industry needs to actually solve the underlying problems. The cascade needs to continue inside labs and also among the media and politics.

SAIR'S OPEN MATH MODEL INITIATIVE

The Foundation for Science and AI Research (SAIR) was founded with some private donors to create a non-profit organization that could support responsible uses of AI in mathematics and the other sciences, independent of the major AI companies. It plans to develop open models with the mathematical community, with participation open across institutions, regions, and career stages. The initiative is still in early stages, but the foundation is now ready to collect expressions of interest. It is also seeking partners who can contribute to funding, compute, expertise, or community-building efforts.

architectural patterns. Get your digital copy

INTRODUCING GROK VOICE TRANSCRIBE 2.0

Grok Voice Transcribe 2.0, launching today, boasts double the accuracy of its predecessor at the same price, excelling in noisy, multilingual environments.

META TO GIVE MUSE ITS OWN MAILBOX FOR COMMUNICATION

Meta plans to introduce a dedicated Mail tab for Muse, creating a separate mailbox for its communication.

STRING-MATCHING EVALS CAN REWARD AI AGENTS FOR FAKE COMPLIANCE

Checking whether generated code contains “Azure” only proves that the word appears somewhere, not that an agent actually used Azure or built working software.

ATTACKERS CAN TURN AN AI AGENT'S OWN TOOLS AGAINST IT

Goal hijacking works when an agent treats retrieved content as instructions, letting a webpage, email, or document redirect its connected tools.

NYSE SPENT A YEAR TESTING AVALANCHE TECHNOLOGY FOR TOKENIZATION PLANS

Ava Labs President Charley Cooper said the NYSE has spent the past year testing Avalanche technology and probing its economics as the exchange builds a platform for onchain trading and settlement of tokenized US equities and ETFs. Cooper described a "close working relationship" with NYSE parent ICE but stopped short of saying Avalanche had been selected, and ICE's Michael Blaugrund said the firm remains "very engaged" in evaluating blockchain platforms. NYSE announced the tokenization platform in January and brought on tZERO as a design partner in August, but has yet to name a settlement chain. Cooper expects some smaller venues to offer 24-hour weekday trading within a year, aided by the SEC's new innovation exemption for onchain tokenized-stock trading.

VISA TO CLOSE CROSSMINT MEMECOIN REWARDS LOOPHOLE

Visa is moving to stop payment processors from categorizing memecoin purchases under merchant code 5815, the "digital media" category that lets credit card buyers earn ordinary points or cash back, following The Block's investigation into Crossmint's memecoin checkout on Fomo and Robinhood Wallet. Checkout.com, Crossmint's processor, was reportedly told directly to end the practice, and Visa's grace period is expected to end next week, after which purchases must be processed under standard crypto transaction rules. Crossmint already pulled tokens GENIUS and DEGEN from its Apple Pay checkout after The Block questioned their classification as memecoins, and New York Attorney General Letitia James' office said it is looking into the matter. Mastercard has not disclosed whether it will follow Visa's lead.

LLM RUNS ENTIRELY ONCHAIN VIA SOLANA SVM

The first LLM inference run executed as pure onchain Solana transactions, with model weights, activations, attention calculations, and token outputs each processed as standard SVM program calls and no off-chain compute involved. The Solana Virtual Machine served as a general-purpose ML compute substrate, making every inference step an auditable, verifiable ledger entry rather than a trusted black-box computation. This removes the off-chain intermediary that underlies most current AI agent and oracle architectures, where inference results are attested after the fact rather than proven in place. If per-inference costs compress to practical levels, the pattern enables a class of trustless on-chain AI agents whose reasoning is natively verifiable by any party reading the chain.

S&P GLOBAL ACQUIRES OPENZEPPELIN

S&P Global agreed to acquire OpenZeppelin, whose smart contract library underpins the majority of major DeFi protocols, stablecoins, and tokenized funds and has processed over $37 trillion in cumulative value transferred. The deal, financial terms undisclosed, extends S&P's ratings, benchmarks, and risk assessment infrastructure into onchain markets as tokenized finance scales toward systemic financial infrastructure. Since its 2015 founding, OpenZeppelin has completed 900+ security engagements and surfaced 10,000+ pre-production vulnerabilities across the ecosystem. OpenZeppelin Contracts and all open-source tooling remain free and publicly maintained on GitHub, with the existing audit and engineering teams continuing operations unchanged.

POST-QUANTUM SECURITY: THE HIDDEN UPGRADE

Google's 2026 quantum computing progress report set a 2029 PQ readiness target, converting the migration from deferred planning to current procurement after NIST finalized ML-KEM (FIPS 203), ML-DSA (FIPS 204), and SLH-DSA (FIPS 205) as binding standards. The "harvest now, decrypt later" attack vector is operational today, with adversaries archiving encrypted data for decryption once fault-tolerant quantum hardware matures, a threat with acute implications for blockchain infrastructure where ECDSA and EdDSA wallet key material and transaction signatures are permanently on-chain and cannot be retroactively re-encrypted. Migration spending is concentrating in four areas: hardware security modules and trusted execution environments, crypto-agile embedded devices, machine identity and certificate lifecycle management, and quantum-safe network deployment across telecom and sovereign infrastructure. Thales, Infineon, CyberArk, SK Telecom, and NTT are the named beneficiaries of this multi-year cycle, while blockchain protocols face the same ECC replacement pressure as traditional PKI infrastructure and cannot address it through a software patch alone.

WHEN EVERY TRADER USES AI, NOBODY BEATS THE MARKET

When all retail traders use the same LLMs with the same basic prompts, models converge on identical trade recommendations, producing herding rather than the efficiency gains classical EMH theory would predict. This convergence creates an adversarial opening: corporate IR teams will calibrate earnings disclosures to exploit model biases, and hedge funds will pre-position against the predictable retail AI consensus before it forms, with momentum strategies then amplifying the resulting price dislocations from fundamentals. The structural edge shifts to actors with proprietary data, superior or faster models, and the capacity to develop non-consensus theses, a gap that widens as retail AI tooling improves but remains generic.

BLOCKCHAIN CAPITAL ALLEGEDLY RUNNING QUIET COUP AT POLYMARKET

There are rumors alleging Blockchain Capital is running a coordinated founder-displacement campaign at Polymarket. Using the same playbook it allegedly executed at Bluesky earlier this year when partner Kinjal Shah took a board seat after leading the Series A, Jay Graber was moved to "Chief Innovation Officer," and Toni Schneider was installed as CEO. At Polymarket parent Blockratize, BCap partners Bart Stephens and Joshua Rivera have secured formal director seats, while new CFO Warren Jenson (also Audit Chair at BCap portfolio company Ripple) has removed founder Shayne Coplan's unilateral control over budgets and capital allocation, and new VP of Ops Collin McKinney Hill (Ray Dalio's former Bridgewater Chief of Staff, ex-DoorDash GM) has ring-fenced day-to-day operations. None of the claims have been independently corroborated, and all named parties have yet to respond publicly.

STABLECOINS ARE TURNING BANKING INTO A ROUTING DECISION

Stablecoins and self-custodial accounts decouple money from any single bank, letting financial providers plug in to compete on specific jobs like FX pricing, settlement speed, or local payment rails rather than owning the full customer relationship. The emerging model is similar to an aggregator for financial services, where switching providers becomes a routing decision instead of moving one's entire financial life. Fierce competition on FX, rails, and compliance is pushing the stablecoin orchestration ecosystem toward more efficiency for customers.

ROBINHOOD CHAIN COLLECTED $4.5 MILLION IN DAILY FEES

Robinhood Chain collected roughly $4.5 million in transaction fees on September 3 while paying Ethereum only about $398 for data posting and proof costs.

KOREAN BROKERAGE, NYSE, AND PAXOS ARE QUIETLY BUILDING ON AVALANCHE

A South Korean brokerage managing $200 billion built a tokenized securities platform on Avalanche, while the NYSE is separately developing its own blockchain trading platform that Avalanche is meeting the technical requirements for.

FIRST RESULTS FROM ETHFI CASHBACK CLAIMS FOR ETHERFI CASH USERS

Since cashback became available Friday, EtherFi Cash users have logged 225 claim transactions totaling 11,145 ETHFI.

ORDERING A PHYSICAL PLASMA CARD

Plasma's card-ordering flow pre-fills the shipping address from onboarding data and uses a press-and-hold confirmation with haptic feedback, but there's a $20 shipping fee against otherwise near-zero card fees.

data pipeline patterns.

TRAINING A 4B MODEL TO PRODUCE 81% FASTER QUERY PLANS THAN POSTGRES

A 4B Qwen model was trained with supervised fine-tuning and reinforcement learning to generate faster PostgreSQL query plans. Across 113 join-heavy queries, its best-of-three results achieved a 1.81x geometric mean speedup over PostgreSQL's default plans and reduced total workload latency by 44.7%, showing small models can learn useful database optimization strategies from execution-time feedback.

THE ARCHITECTURE OF NEKI

PlanetScale's Neki horizontally shards standard PostgreSQL while presenting applications with a single Postgres connection, using routers, sidecars, replication, and automated failover to manage distributed queries and availability. It can also reshard data, move tables, and perform online schema changes without taking the database offline.

SAVING ANOTHER 100TB OF RAM WITH MATH (AND RUST)

Consistent hashing smooths load by assigning each server many virtual points, but those points can become a major memory cost at global scale. Cloudflare reworked Pingora's hash-ring representation, sorting, and rollout path to reclaim more than 100 TB of RAM. The article connects the distribution math to a staged migration that contained cache churn and preserved rollback control.

DEBUGGING INCONSISTENT QUERY LATENCY ON A POSTGRESQL HYPERTABLE: WHAT WE LEARNED

A PostgreSQL hypertable query was fast by hand but intermittently stalled because JDBC prepared statements switched to generic plans. Without parameter values, the planner could not exclude old chunks and spent seconds planning instead of executing. Logging the bind phase exposed the fault, while disabling server-side prepared statements and reducing chunk count restored predictable latency.

TWO TECHNIQUES FOR WORKING WITH SYSTEM ONE MODELS

System One models act as fast, general-purpose classifiers that make many simple decisions in parallel rather than generating long responses. Two useful patterns are to give them layered goals for real-time tasks and use tournament-style comparisons when choosing the best option from a large set.

THE SPAN ATTRIBUTE THAT BLEW UP YOUR OBSERVABILITY BILL

A high-cardinality span attribute can quietly multiply metric series even when trace volume and sampling remain flat. Sampling reduces stored trace bytes, not the distinct label combinations produced upstream by span metrics. Keep identifiers on traces, bucket them before metrics aggregation, and set cardinality and expiration limits so one deployment cannot turn a temporary dimension into a lasting observability bill.

SATURATION AT GITHUB: THE SAGA CONTINUES

A GitHub outage was triggered when a background cleanup job saturated a shared database primary, while monitoring incorrectly showed the cluster as healthy because it only watched replica lag. Long request timeouts and repeated retries then amplified the failure, highlighting how partial health signals and retry storms can turn database saturation into a site-wide incident.

together →

PG-JEV

pg-jev brings natural-language classification directly into PostgreSQL, letting you filter, rank, and classify rows with Jev-powered SQL functions that return calibrated probabilities instead of generated text. It batches and caches evaluations to make semantic queries practical without embeddings or vector indexes.

CHARTS BUILT FOR CHAT

dbt Charts is an open-source, YAML-based dashboard language designed for humans and AI agents to build governed charts in code rather than through a BI interface. Dashboards stay auditable and version-controlled alongside dbt models, with validation catching broken queries and poor visualisations before they ship.

FASTER JSON PARSING WITH SVE2 ON ARM PROCESSORS

JSON ingestion often spends significant time locating structural characters before parsing values. A new simdjson implementation uses ARM SVE2's match instruction to classify those characters more directly, improving indexing throughput by 3-9% on Graviton 4 and 5. The full parser gain is smaller, but the patch shows how targeted SIMD changes can improve a heavily optimized data path.

JEV FOR 10-K DATA EXTRACTION: FAST & CALIBRATED

Jev can act as a fast, inexpensive verification layer for extracting structured data from 10-K filings: another tool finds candidate values, then Jev selects the correct one and returns calibrated confidence. This avoids unreliable generate-and-parse loops while routing ambiguous cases to larger models or humans.

ADDING AN INDEX MADE THIS QUERY SLOWER

An index matching ORDER BY in Postgres can still hurt when the filtered rows sit far down that index: inspect Rows Removed by Filter, test representative parameters, and use a composite index that matches both the filter and ordering.

PGRUN

PGRun gives agents and CI fast, isolated Postgres branches from production, avoiding shared staging conflicts, and keeping production protected.

GOOGLE'S NEW ‘CC' IS AN AI AGENT THAT HELPS FAMILIES RUN THEIR HOUSEHOLDS

Google is testing a family-focused version of CC that turns shared emails and messages into daily briefings, calendar events, and tasks for up to six household members. The agent has its own Google account and can help fill out permission slips, create shopping lists, and plan meals, while each person controls what information they share. The experiment is available through a waitlist for US adults with personal Google accounts.

ANTHROPIC TRIES TO MAKE CLAUDE STICKIER WITH LAUNCH OF DOCS AND SLIDES

Anthropic launched Claude Docs and Slides, a rich-text editor letting users draft documents and presentations from chat while Claude asks clarifying questions and comments on its choices. Files land in the Artifacts tab, export to Word, PDF, Google Docs, or Markdown, and support real-time collaboration, though Team and Enterprise still lack version history. Analysts see it as an attempt to make Claude stickier by pulling productivity work into the AI, unlike Microsoft and Google embedding AI into existing office tools.

DISNEY'S FIRST CTO LED AN AI STARTUP IT ONCE ACCUSED OF COPYING ITS CHARACTERS

Disney has appointed former Character.AI CEO Karandeep Anand as its first chief technology officer, reporting to CEO Josh D'Amaro. The hire comes after Disney accused Character.AI of using its copyrighted characters without permission and demanded their removal. Anand previously held leadership roles at Meta and spent 15 years at Microsoft, bringing consumer technology and AI experience to Disney's push to make technology a larger part of its entertainment business.