Search the archive

PYSCN

pyscn is a code quality analyzer for Python vibe coders that performs structural analysis to keep code bases maintainable.

HOW I INFLUENCE TECH COMPANY POLITICS AS A STAFF SOFTWARE ENGINEER

Let executives set the overall priorities for the company and tailor your own technical plans to fit.

HOW ONE AI MODEL CREATES A PHYSICAL INTUITION OF ITS ENVIRONMENT

Video Joint Embedding Predictive Architecture (V-JEPA), a model created by Meta, uses ordinary videos to understand the physics of the real world.

MEMORY ACCESS IS O(N^[1/3])

If you can break up a task into many parts, each of which is highly local, then memory access in each part will be O(1), but if the task requires a lot of memory interdependencies, then you will get lots of O(N^⅓) terms.

HOW GDP HIDES INDUSTRIAL DECLINE

GDP is a very complicated statistical construct made by government bureaucrats - the public doesn't really have a way to replicate, audit, or verify assumptions.

PAY YOURSELF FIRST

Do the work that actually means something to you first, or you'll never get through the endless list of work that could be done.

OPENAI PREPARES TO RELEASE AGENT BUILDER DURING DEVDAY ON OCTOBER 6

OpenAI's Agent Builder will help users build agentic workflows and connect MCPs, ChatKit widgets, and other tools. It is a direct competitor to established workflow automation tools like n8n and Zapier. The Agent Builder features a drag-and-drop canvas that allows users to create agent flows from predefined templates. The canvas supports a range of modular building blocks, nodes for logic, connectors, and more, mirroring the flexibility seen in other agentic workflow platforms.

GEMINI'S HOME SCREEN COULD SOON GET A DISCOVERY-STYLE REDESIGN

Google appears to be testing a new home screen for Gemini. The redesign adds a scrollable feed of one-tap prompt suggestions, with prompts ranging from image edits to requests regarding news, quizzes, and coding. Video and screenshots of the new home screen are available in the article. There is no guarantee that the redesign will roll out broadly.

WITH ITS LATEST ACQUI-HIRE, OPENAI IS DOUBLING DOWN ON PERSONALIZED CONSUMER AI

OpenAI has acquired Roi, an AI-powered personal finance app. Only the company's CEO will be joining OpenAI. Roi will wind down operations and end its service to customers on October 15. Terms of the deal have not been disclosed. The acquisition clearly aligns with OpenAI's bet on personalization and life management as the next layer of AI products.

OPENAI'S FIRST DEVICE WITH JONY IVE COULD BE DELAYED DUE TO 'TECHNICAL ISSUES'

OpenAI and Jony Ive's partnership is struggling with some technical issues that could end up pushing back their highly anticipated AI device's release date. OpenAI and Ive are still working on the assistant's personality and potential privacy concerns. The project's budget could reportedly be challenged due to the increased computing power necessary to run mass-produced AI devices. There is still little information about the upcoming product publicly available.

WHICH TABLE FORMAT DO LLMS UNDERSTAND BEST? (RESULTS FOR 11 FORMATS)

Understanding format sensitivity in large language models is crucial for data pipeline architecture, performance optimization, and cost management. Many RAG pipelines involve ingesting documents that contain tables of data. These tables need to be formatted in a way that is easy for models to consume otherwise they may be needlessly hurting the accuracy of the overall system. Markdown-KV appears to be a good default to use in situations where accuracy is paramount. CSV and JSONL can hurt system accuracy.

CIRCULAR FINANCING: DOES NVIDIA'S $110B BET ECHO THE TELECOM BUBBLE?

Analysts immediately drew comparisons to the telecom bubble on the news that Nvidia had invested $100 billion in OpenAI in September. During the telecom bubble, demand was speculative, and customers burned cash. Many of Nvidia's customers are profitable and sophisticated hyperscalers. This suggests that the current situation is different from the telecom bubble.

ECONOMICS AND AI

Most economists refuse to discuss the possibility of transformative AI, even at workshops designed for it, insisting that comparative advantage keeps human wages high, as past tech progress has shown. AI's impact depends on domain structure, like protein folding has predictable patterns, but mapping genomes requires individual observations.

AI CYBER DEFENDERS IN PRACTICE

Claude Sonnet 4.5 has been enhanced with strong cybersecurity capabilities, matching or surpassing previous frontier models in tasks like vulnerability detection. This shift reflects a growing focus on using AI not only for cyber offense but also to assist defenders in securing systems.

👉JOIN THE LIVESTREAM

JULES API

The Jules API gives developers programmatic access to Jules' capabilities to automate and enhance software development cycles. It can be used to create custom workflows, automate tasks like bug fixing and code reviews, and embed Jules' intelligence directly into everyday tools. This page provides a quick overview of the API and walks through how to make an API call.

ANATOMY OF A MODERN FINETUNING API

Mira Murati's AI lab, Thinking Machines, has released a language model fine-tuning API called Tinker. Tinker exposes a handle of very low-level functions to users to let them perform supervised fine-tuning and online reinforcement learning. Each training step requires sending a batch of data over the network. It offers a stable, mature infrastructure for RL and fine-tuning experiments, making it possible for anyone to be an AI researcher.

SINQ

SINQ (Sinkhorn-Normalized Quantization) is a novel, fast, and high-quality quantization method designed to make any large language models smaller while keeping their accuracy almost intact. It uses dual scaling to make models less vulnerable to outliers. SINQ's Sinkhorn-normalized optimization, which iteratively rescales rows and columns to balance their variance, makes errors more spread out and less severe, preserving model accuracy even at 3-bit precision. Weights become inherently easier to quantize when the overall matrix imbalance is reduced, leading to more stable behavior across layers and consistently higher accuracy even at very low bit-widths.

IT'S A JAX, JAX, JAX, JAX WORLD

Stan's days are likely numbered. High-end applications are replacing Stan with JAX. The biggest obstacle for the shift is finding hardware on which to run JAX most effectively, but this won't be a problem in ten years. The Stan project will continue development, as it will likely still be used for a long time. The team has plenty of strategies for making it faster, like adding samplers that will work well on CPU but not on GPU.

ON-DEVICE OCR WITH CORE ML AND MLX

dots.ocr was adapted to run entirely on-device using Apple's Core ML and MLX frameworks. In this post, Hugging Face outlines the conversion process, offering a practical guide for developers looking to deploy similar models locally.

2 HOURS IN LINE FOR A FREE HAT

The Claude team recently gave out hats and coffee at an event where people had to wait two hours in line. This post details how the event went down. Participants at the event waited, took pictures, posted on social media, and became one of the cool kids. The Claude team created an experience people wanted to be part of, which made it worth waiting for hours just for a hat. Pictures of the event and the merch the Claude team handed out are available in the article.

THIS META ALUM HAS SPENT 10 MONTHS LEADING OPENAI'S NATIONWIDE HUNT FOR ITS STARGATE DATA CENTERS

Keith Heyde, OpenAI's Head of Infrastructure, leads site development within OpenAI's industrial compute team, a division that's quickly becoming one of the most important groups in the company. Infrastructure is now a strategic pillar on par with product and model development. OpenAI is betting that owning the next generation of physical infrastructure is central to controlling the future of AI. This article looks at how Heyde has spent the past several months searching for Stargate data center sites and the challenges that face the project.

AI CHIPMAKER CEREBRAS WITHDRAWS IPO

The company offered no explanation for its decision, but the move came days after raising $1.1 billion from private markets.

HOW CALIFORNIA STRUCK A LANDMARK LAW TO REIN IN ROGUE AI

Google, Meta, and Amazon operated as a bloc pushing for higher revenue thresholds to limit which companies would be covered, and a16z and OpenAI suggested waiting for federal frameworks.

SOME GUY TRAINED GPT 1 THINKING

GPT-1-thinking performs exceptionally poorly compared to today's cutting-edge models.

OPENAI'S HUNGER FOR COMPUTING POWER HAS SAM ALTMAN DASHING AROUND THE GLOBE

Sam Altman is looking to East Asia and the Mideast for funding and manufacturing partners for his plan to build more data centers.

BITCOIN BREAKS $125K AMID “UPTOBER” MOMENTUM

Bitcoin set a new all-time high of $125,700 on Sunday, surpassing its August record as “Uptober” gains continued despite a US government shutdown. Analysts attribute the rally to an accumulation phase, cooling speculative activity, and easing long-term holder sell pressure, with spot Bitcoin ETFs seeing their second-highest weekly inflows since launch.

COINBASE SEEKS OCC CHARTER TO BRIDGE CRYPTO AND TRADFI

Coinbase is pursuing a national trust company charter from the US Office of the Comptroller of the Currency (OCC), joining firms like Paxos, BitGo, Ripple, and Circle in applying for federal licensing. The charter would streamline regulatory oversight and support Coinbase's goal of integrating digital assets more deeply into traditional finance while maintaining strict compliance and security. Coinbase has no intention of becoming a bank. The move is a step toward clear federal rules that enable responsible innovation.

METAMASK CONFIRMS $30M ONCHAIN REWARDS PROGRAM

MetaMask's upcoming Rewards program will feature referral bonuses, mUSD incentives, partner rewards, and access to future tokens. It is not a yield farming scheme. Season 1 alone will distribute over $30 million in LINEA tokens, making it one of the largest onchain reward programs ever. Long-time users will receive exclusive perks tied to the future MetaMask token, with full rollout details coming in the next few weeks.

SAMSUNG WALLET PARTNERS WITH COINBASE ONE

Samsung Wallet has expanded its Coinbase partnership to offer Galaxy users special access to a free three-month subscription to Coinbase One, providing boosted staking rewards, exclusive partner offers, and zero trading fees on select assets for US users subscribing to the Standard Tier. The enhanced integration builds on Samsung's July partnership, which brought full Coinbase cryptocurrency payments and management to Galaxy devices, initially rolling out to select users in the US and Canada in August with plans for broader expansion. New Coinbase users making their first trade through Samsung Wallet receive a $25 credit, while the premium membership experience aims to streamline crypto access. Samsung continues its push into convenience-friendly payments, complementing recent Samsung Wallet features, including NFC-based peer-to-peer payments between devices in the same room and the ability to split in-store payments into installments for easier transaction management.

SUPPORTIVE SIGNALS, GOLD AND STABLECOINS

Coinbase maintains a tactically bullish October outlook driven by US dollar softness, short-term global liquidity impulse, and Fed policy bias toward insurance cuts, with the government shutdown potentially delaying key economic data releases and forcing markets to rely on private gauges like ADP employment showing essentially zero net job creation. SWIFT's announcement to embed a blockchain-based shared ledger into its infrastructure for bank-grade stablecoin settlement represents a watershed moment that formally mainstreams the digital economy, coinciding with adjusted stablecoin transaction volumes reaching $3.07 trillion in September and surpassing major retail payment rails. Tether is reportedly seeking $15-20 billion in fresh capital for approximately 3% equity at a potential $500 billion valuation, with USDT's dominant 58% market share reinforced by a surge in perpetual futures activity that predominantly quotes and margins in USDT, driving mechanical circulation increases. Gold is not a reliable bitcoin indicator due to inconsistent correlations swinging from -0.8 to +0.8 since 2013. Global liquidity is the most reliable macro signal, with consistent ~0.9 correlations over the past three years. Spot bitcoin ETFs attracted $3.5 billion in September inflows, comparable to gold ETFs' $4.2 billion despite bitcoin's recent underperformance.

CRYPTO'S ENDGAME: PROOFS AND PRAGMATISM

Zero-knowledge proofs excel in specific contexts like rollup security and high-stakes nation-state protection but are overkill for everyday transactions, with proof generation costing around ten cents per simple action versus near-free signatures, leading to adoption primarily as invisible infrastructure rather than user-demanded features. ZKML compounds this problem by stacking GPU-intensive proof generation on top of GPU-intensive inference, creating costly solutions seeking problems to solve, while fully homomorphic encryption still depends on MPC/TSS for distributed key control to avoid centralized trust points. Smart accounts face fundamental limitations by trapping automation inside single blockchain state machines, unable to enforce simple cross-chain rules like daily spending limits that span multiple networks and contexts where real user identities operate. The pragmatic alternative uses threshold cryptography with MPC/TSS+TEE as a layered model where smart accounts handle narrow on-chain guardrails while threshold systems enforce sophisticated cross-chain policies, multi-context identity, and higher-order automation above individual blockchains.

ETHREX: MINIMALIST L2 WITHOUT THE TOKEN PROBLEM

Layer-2 stacks are structurally unsustainable because their tokens undermine long-term developer incentives. Once vesting ends, core teams lose motivation, leaving bloated, hard-to-maintain codebases behind. Ethrex avoids this trap with a minimal, maintainable design that functions as both an L1 and L2 client and carries no token, enabling anyone to launch an L2 with a single command. Its open, tokenless model and forthcoming documentation aim to make Ethereum scaling infrastructure fully self-sustaining and community-driven.

POLYGON TOKEN HOLDERS' MANIFESTO

Polygon leadership needs to act to reverse $POL's severe underperformance (down 46% year-over-year and 92% below its ATH) despite a strong market. Persistent 2% inflation, unclear execution around AggLayer, and waning holder confidence indicate that $POL risks becoming a “dino token” without decisive intervention. The team should end inflation, introduce buybacks or burns, and restore transparency to rebuild investor trust and spark a recovery toward $0.71 as a first sign of reversal.

ARGENTINA'S ETHEREUM AFFECTION

Argentina has an unusually deep grassroots awareness of Ethereum.

BEST CRYPTO PAYMENT RAILS SNAPSHOT

Polygon leads in Stripe stablecoin payments with a broad ecosystem, Plasma's purpose-built L1 has a $5.5B stablecoin supply and $6.2B in Aave deposits in its first week, Solana has actively pushed P2P payments via Solana Pay and card partners, and Base has Coinbase-backed growth with USDC payments on Stripe and Shopify.

ANDRE CRONJE'S NEW CRYPTO PROJECT

DeFi pioneer and creator of Yearn Finance, Andre Cronje, raised $200 million at a $1 billion valuation for his new project, Flying Tulip.

FIRST TOKENIZED INDEX FUND

Centrifuge's SPXA is the first tokenized S&P 500 Index Fund onchain.

GOOGLE'S AI MODE GETS BETTER AT UNDERSTANDING VISUAL PROMPTS

Google's AI Mode now better understands visual search prompts through an enhanced "query fan-out" technique, which runs multiple background searches to interpret complex requests, such as "moody but maximalist" bedrooms. The update makes the chatbot multimodal, allowing users to initiate conversations with images or videos and receive more visual responses rather than text-heavy answers. These improvements particularly benefit shopping searches, enabling AI Mode to handle nuanced queries such as "barrel jeans that aren't too baggy" with refinable follow-up questions.

APPLE GLASSES ARE COMING: HERE'S EVERY RUMORED NEW FEATURE

Apple is developing AI-driven Apple Glasses expected in 2026, with release likely in 2027. The first version will lack a display or AR features, focusing instead on visual intelligence, Siri voice control, multiple cameras, health tracking, and deep iPhone integration. Powered by a low-power chip and built-in speakers, the glasses will function as an iPhone accessory, with a future model reportedly planned to include a display.

META PLANS TO SELL TARGETED ADS BASED ON DATA IN YOUR AI CHATS

Meta will use data from user interactions with its AI products—including Meta AI chatbot conversations, Ray-Ban smart glasses recordings, and AI-generated content—to sell targeted ads on Facebook and Instagram starting December 16. The policy applies globally except in South Korea, the UK, and the EU, with no opt-out available for affected users. While Meta claims conversations about sensitive topics won't influence ads, the company acknowledges that it's still developing systems to leverage this AI data for its advertising business.

WHERE'S THE AI DESIGN RENAISSANCE?

Despite 2.5 years of intense AI hype, there's no evidence that current AI tools are making design faster, eliminating designer jobs, or enabling the promised software renaissance—metrics like new app releases and domain registrations remain flat. AI struggles with designs that require tight constraints, novelty, or complexity because its architecture is based on training data, making it best suited for average designs rather than bold, opinionated work that stands out. Designers should focus on out-of-training-data skills, such as pixel-perfect UI and novel interactions, while using AI primarily for small personal projects, internal prototypes, and simple public tools where security and complexity are minimal.

BORING IS GOOD

AI language models should focus on "boring" tasks, like proofreading and summarizing, rather than attempting high-level thinking, where hallucinations and faulty reasoning can lead to failures. Small language models (SLMs) are sufficient for these modest language tasks, offering benefits such as easier ethical training and lower costs. The industry has been using LLMs backwards by trying to automate complex tasks from the top down instead of augmenting simpler tasks from the bottom up.

HOW DO GOOD DESIGNERS GET STUCK? THEY WASTE TIME TRYING TO MIND-READ THEIR TEAM

Good designers often get stuck by attempting to mind-read their team's intentions when designs are rejected, creating false narratives like "they don't value UX" instead of addressing the real issue. This "LoserThink" pattern prevents designers from asking whether they're solving the actual problem their stakeholders need solved. The solution is to replace assumptions with direct questions about success criteria and understanding team pressures through brief conversations.

WAN 2.5 AI VIDEO GENERATOR

Describe what you want to see, and Wan 2.5 will bring it to life with stunning visuals and perfectly synchronized audio.

YOUR SPACE FOR NOTES, TASKS, AND BIG IDEAS

Create stunning documents, manage tasks, build dynamic collections, and structure your ideas your way, whether it's personal journals, work projects, or daily planning.