Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
Sponsored
After passing 15,000+ GitHub stars, how does Google keep open-source agent skills reliable at scale? READ THE ENGINEERING BREAKDOWN [13] on designing skills with remote MCP tools, ADK authoring loops, and continuous evals for accuracy and token efficiency. Building your own agent stack? Register for HARNESS ENGINEERING: CRUCIAL SCAFFOLDING
Priming is a technique that involves using pretrained Transformer weights to build models with more efficient architectures at less than 1% of the original training cost and with throughputs up to twice as high. It allows researchers to explore the design space, the layer type, the hybrid ratio, and layer placement cheaply and systematically. Every variant starts from the same strong initialization and requires minimal training compared to pretraining. Priming is not tied to a specific Transformer family or model class, and extending it to other models requires no algorithmic changes.
On-policy distillation has become increasingly popular for transferring teacher capabilities to student models. Semi-OPD, which distills from offline rollouts generated by the initial student, can often outperform OPD in both accuracy and training efficiency. The choice between OPD and Semi-OPD depends on the alignment between the initial teacher and student. OPD is beneficial only when the two are highly aligned with high overlap ratios.
Legal practice demands an unbroken chain between facts, statutory doctrine, analytical reasoning, and client-ready work product. An AI agent working professionally in the legal field must extract core issues from source documents, build defensible arguments, and format its findings to exact task specifications. Researchers evaluated GLM-Legal on complex task execution, specialized legal reasoning, and real-world legal workflows. It achieved a 19.2% all-pass rate on Harvey LAB while matching frontier closed-source models on LegalBench and PLawBench despite running on a significantly leaner parameter footprint.
Several Claude users noticed over the weekend that Claude Opus 5.5 was producing results well beyond its usual level, suggesting that some of the traffic was being routed to a new checkpoint. Examples of outputs from the new model are available in the article. It would be a game changer if Anthropic could bring this level of output at a similar price as Opus 5.5. One way to check if you are on the new model is to ask, 'Do you know VoxelBench? Do not search' - if it answers, you're probably on the new model.
Sponsored
Deploy a RAG pipeline to your AWS with one Terraform template. Sign up for Elastic Cloud free [19], upload documents to S3, then query
The model is being rolled out as a gray release to randomly selected subscribers.
Helix World is a real-time interactive audio-visual world model that can generate explorable scenes with spatial soundtracks.
Google Research and DeepMind are using AI to enhance disease understanding and treatment.
Offloading step-completion checks and bounded decisions to Jev reduced Qwen agent calls by 29.6%, estimated API costs by 26.6%, and latency by 19.7%.
AGENT-ENGINEER is an open-source plugin that enhances AI coding assistants, offering 11 skills, 4 subagents, and 5 hooks to tackle agent-engineering challenges.
Amy Oldenburg, Morgan Stanley's head of digital asset strategy, said the bank now expects to support tokenized equities on its alternative trading system in Q1 2027, later than the H2 2026 timeline it gave in March, because client demand depends on products that do not exist yet. She pointed to DTCC's tokenization service launch at the end of October and the GENIUS Act taking effect in January as milestones, and said the tokenized RWA market will “without question” land in the $3 trillion to $10 trillion range by 2030. The bank is also building abstracted wallet infrastructure, runs a Digital Asset Lab, and holds a conditional OCC approval for a digital trust.
Caroline Pham, now MoonPay's chief legal officer, said the CFTC does not need Congress to regulate crypto and called its proposed Regulation CAM and CTX the “absolutely the kind of protection” retail crypto traders need. She argued tokenized deposits, not stablecoins, fit wholesale markets because atomic settlement of one trading day would need about $15 trillion in cash that does not exist, and expects NYSE and Nasdaq to list “real” tokenized shares with meaningful volume within two years.
The Stonks Exchange acquired Stockify, combining Base's primary token launch venue ($STONKEX) with a protocol that auto-distributes tokenized equity dividends via a no-claim mechanism. Under $STFY's unchanged mechanics, a 3% ETH trading fee routes 90% of proceeds to purchase tokenized stocks sent to eligible wallets with no staking or claiming required. The deal adds $STONKEX to Stockify's dividend index at equal weight alongside NVIDIA, Meta, Apple, Alphabet, and SpaceX, giving STFY holders direct exposure to the combined entity's growth through existing payout infrastructure. Base now hosts 130+ tokenized stocks targeting 250 by month's end, and the merged team aims to supply liquidity and holder depth for newly listed assets as the chain accelerates its push to become the primary onchain equities venue.
AMM liquidity providers on Ethereum lose money to LVR (loss versus rebalancing), with MEV bots capturing most arbitrage profits from price gaps between pools and external markets. BRO-AMM addresses this by combining oracle-verified post-trade reconciliation windows, mandatory trader collateral, and reputation-tiered fees to compensate legitimate arbitrageurs while penalizing excess extraction. Six-month ETH/USDC simulations show 68-79% reductions in arbitrage losses on Ethereum and 49-55% on Arbitrum, with Ethereum depositors shifting from net losses to profitability. The proposal depends on trusted oracle data and remains simulation-only, pending live validation.
Two wallets are exploiting Paper Trade by running $20M clips on Hyperliquid to shift ETH prices 10-20bps, then opening nine-figure paper longs that capture the manufactured move at no real cost. The attack exposes a structural gap between real-market price impact and paper trade settlement that the protocol must close to remain viable.
EtherFi introduced its own stablecoin, ether.fi USD, powered by Ethena, and picked MoonPay as a payments infrastructure partner; RedotPay launched SPEI for Mexican pesos and partnered with Niu Global; Stripe Treasury made OUSD on Tempo the default stablecoin for US businesses, with OUSD-backed cards paying 2% cash rewards; Tangem Pay became a physical card (5,000-card early run); and Jupiter Mobile launched on-chain QR payments.
The IMF's Global Financial Stability Report found that more than half of tokenized equity trading occurs outside US market hours, with 80% of transactions in the five most liquid tokenized US stocks (Tesla, Nvidia, and Google among them) sized below one share. These tokens carry roughly 1.5x the volatility of their traditional counterparts while showing considerably lower liquidity, raising structural questions about price formation and investor protection. The tokenized equity segment totals $2.3 billion within a broader $65 billion tokenized real-world asset market, and the IMF data positions the segment as a retail-driven, offshore-hours venue rather than a functional extension of regulated equity markets.
Raheim Hamilton, co-creator of Empire Market, received a 40-year federal prison sentence and a $5 million fine on October 9 for running a dark web marketplace that processed over 4 million transactions totaling $430 million between February 2018 and August 2020. The platform accepted only cryptocurrency and encouraged vendors to use tumbling and mixing services, with drug sales accounting for roughly $375 million of volume. Prosecutors seized approximately 1,230 Bitcoin and 24.4 Ether from Hamilton. Co-creator Thomas Pavey faces forfeiture of roughly 1,584 Bitcoin, gold bars, and real estate before his own sentencing later in October. The platform shut down in August 2020 in an alleged $30 million exit scam, and the combined recovery of 2,800-plus Bitcoin across both defendants demonstrates that chain analysis pierced mixing-service obfuscation at this scale.
Payward received full market approval from Dubai's VARA, covering retail, professional, and institutional services, and plans to launch Kraken in the UAE within weeks with AED spot and margin trading, direct AED deposits and withdrawals, onchain staking, and xStocks tokenized stocks, plus OTC and Kraken Prime for institutions.
Jeonbuk Bank started offering bitcoin-backed loans on a limited basis to corporate clients, accepting only bitcoin as collateral for now, with plans to expand depending on demand and regulatory developments.
AAVE's share of major lending tokens' combined FDV fell from nearly 60% in January to about 30% in June and is now back at 46.8%, a gain of more than 17 points.
Google's new web-based SynthID Detector opens AI-content checks to anyone who signs in, flagging images, video, and audio from Google, OpenAI, Nvidia, or Kakao models. It misses AI content without a SynthID watermark, such as output from Claude, Grok, and Chinese open-weights models, and cannot tell fully generated media from AI-edited media. Without a universal watermarking standard across all major model makers, a reliable way to detect AI content likely will never exist.
Smoother collaboration and navigation arrive in Figma's mobile app through a page dropdown, a frame preview slider, and share links to the exact page or frame. Individual frames and images export to the phone with format and size options, and the app reopens exactly where users left off.
Apple's second-generation iPhone Air is reportedly coming in spring 2027 with upgrades aimed at addressing several compromises of the ultra-thin original. A second rear ultra-wide camera will join the main camera, while a larger 3,500mAh battery and more efficient 2nm A20 Pro chip should improve battery life. Apple is also expected to add vapor chamber cooling, allowing the thin device to stay cooler and sustain performance for longer.
A year of trying to be anything other than a designer comes as the AI era proves unkind to human-centered designers, especially outside massive companies. Every company has people who absorb organizational mess, work that pays poorly, carries no title, and is sometimes called glue work. The answer is not another heroic generalist but building enough capability into teams to understand customer need, experiment technically, and stop solving problems twice.
AI companies are increasingly turning abstract technology into approachable characters, with mascots such as Muse, Dots, Claw'd, and Notion's animated face giving unfamiliar systems personalities people can understand and relate to. Anthropomorphism satisfies human desires for connection and predictability, while animation can turn a few simple shapes into a character that appears to think, hesitate, search, greet, and react. The strongest mascots become storytelling systems rather than logos with arms, requiring human skills such as observation, empathy, timing, humor, restraint, and a coherent sense of personality.
Products often converge toward similar forms as their purpose becomes clearer, technologies mature, and physical constraints, economics, standards, and competition narrow the range of practical designs. Software can converge even faster, as demonstrated by the rapid spread of chat interfaces following ChatGPT, while emerging AI hardware remains in an experimental phase of pins, glasses, earbuds, and pendants. Whether through functional optimization, market pressures, shared constraints, or simple imitation, successful forms tend to become increasingly alike until a change in purpose or technology opens new possibilities.
Chat with AI to create pages, publish content, and understand your website traffic.
CAKIE adds a transcript and screenshot highlights to your screen recordings so your AI can follow along.
Contralabs surveyed 150k+ working creatives to explore how AI is changing creative work.
TypeSafe's Jev answers choice, score, and yes-or-no questions with a probability for each option instead of writing text. Its training rewards honest probabilities over confident-sounding answers, giving designers the calibrated uncertainty design work needs, though it reads only text and takes questions literally. A new demo adds a maybe zone to HeyStefan's emoji sorter, where Jev parks undecided emoji from a pile of 166.
French illustrator Louise Laborie uses watercolor and charcoal to depict America's motels, highways, skate parks, and other transitional spaces at dusk, populating them with teenagers caught between childhood and adulthood to create quiet, cinematic scenes shaped by loneliness, music, memory, and imagined stories.
CriS Cian turns photo manipulation and digital painting into peaceful, poetic cosmic scenes of futuristic expeditions and encounters beyond Earth.
2026 became the year of the canceled ad, with Pepsi, Novig, Oxford United, Converse, and a golf brand facing backlash, some by accident and some by design.
Sponsored
Over 1M developers use the open-source LiveKit [5]Agents framework (Python or TypeScript) to build voice agents. In less than 10mins, you (or your coding agent) can get a voice agent built and deployed to the cloud for free. Included for free: * VOICE AI TOKENS: one-line swap between models from OpenAI, Google, xAI, and more * A PHONE NUMBER: get your agent to take phone calls in seconds * ONE-COMMAND DEPLOY: ship to LiveKit Cloud from the CLI LiveKit is the dev platform for real-time voice, video, and robotics, powering billions of minutes for companies like Spotify, Headspace, and SpaceXAI. TRY IT FOR FREE
React Compiler uses cache slots and tracks values that change together to avoid unnecessary work. In a 2,000-row microbenchmark, inlining alone slowed rendering from 444 ms to 1,038 ms by removing per-row caches, while adding block DOM updates cut it to 164 ms. These prototypes illustrate optimization opportunities, with workload and memory tradeoffs still to evaluate.
Google's DevSite stored an OAuth secret inside a readable Flask session cookie and reused it as the cookie's signing key, allowing forged sessions to reach unsafe pickle deserialization. Timing-based tests confirmed unauthenticated code execution across the shared platform, earning a $90,000 bounty. The fix separates signing keys from OAuth credentials and removes sensitive serialized objects from client-controlled sessions.
Perplexity Projects uses a lightweight Git forge backed by S3 storage so people and agents can build on durable files from isolated, temporary sandboxes. Submissions must match the latest accepted commit, with stale writers replaying their changes rather than overwriting concurrent work. In production, fewer than 2% of submissions are rejected as stale, with median submit latency below 1.7 seconds.
PostHog's lessons from ten AI products emphasize useful jobs, reliable execution, rich context, and interfaces that let agents do everything users can. Its MCP server has five times as many daily users as PostHog AI, while the install wizard compresses roughly two hours of setup into eight minutes. The advice favors shared infrastructure, observable tool calls, small targeted evals, and UI for reviewing and undoing agent actions.
AI code review will increasingly replace reading every diff, shifting human work toward requirements, acceptance testing, and ownership of releases. Independent review models, predefined guardrails, and automated review-and-fix cycles form the proposed safety net. A trading-bot rounding bug illustrates the tradeoff: fast fixes do not erase losses or compensate for gaps in the specification.
WebMCP lets websites expose JavaScript tools to agents, reducing the slow, brittle interaction of navigating through screenshots and DOM inspection while keeping users in a shared browser session. Shopify's checkout experiment reported 2.7 times faster execution, 58% lower cost, and 60 successful attempts out of 60 compared with 56 for browser automation. Permissions, human approval signals, and testing against complete agent systems remain open challenges.
Sponsored
AI coding tools _haven't_ transformed software delivery... yet. Messy handoffs are largely to blame. In tomorrow's webcast [12] and the follow-up lab on Thursday, see how GitLab Duo Agent Platform triages security findings, filters out false positives, and proposes remediation merge requests, all with human approval built in. Save
Meta introduces Muse Wonder and Muse Wonder Full for coding and other long-running tasks, combining Active Memory for retrieving relevant context with Active Focus for sustained execution. Meta reports stronger performance on unfamiliar tasks from larger base models and plans a fuller safety report.
Open Intelligent UI is an open-source chat interface built with CopilotKit and AG-UI that turns questions into explanations, comparisons, and interactive tools. Its agent chooses text, native components, or custom HTML, CSS, and JavaScript rendered in an isolated iframe, with Jev routing tables to A2UI and richer visuals to Open Generative UI.
Explicit, repetitive Ruby code can make algorithms easy to inspect and resemble the optimized output of ERB and Papercraft template compilers. A proposed quote/unquote mechanism would let Ruby generate code by composing immutable syntax trees instead of concatenating source strings, with examples covering memoization, templates, and routing.
Anthropic demonstrates a 26B model running across 15 inexpensive Android TV boxes and a 9B model setting up inference on eight of nine tested hardware types. GLM-5.2 completed replication in a contained fleet with deliberately trivial access controls, making the result an upper bound on real-world success. Smaller models still struggle to chain the steps together, while projected replication on low-powered devices would take days.
Vals AI tested politically sensitive questions across languages and contexts, finding that Chinese models sometimes criticize China's government while non-Chinese models sometimes echo official narratives. Responses generally leaned more pro-China in Simplified Chinese, while deployment prompts, response formats, and agent harnesses could substantially change the results.