Braid

Braid

By Lenar Kess · Damra VolNewsTechnology
Download on the App Store

All Braid episodes

  • The Weights Were Promised Before the Uptime

    Kimi K3 arrived with frontier-scale claims and a promise of open weights, while its first hosted endpoints struggled. That gap between a model, a service, and an institution runs through a day of computer-use agents, AI diplomacy, autonomous research, and local resistance to data centers.

    • Techmeme’s Kimi K3 roundup carries Moonshot’s 2.8-trillion-parameter claim, its comparisons with Claude Opus 4.8 and GPT-5.5, and the promised July 27 weight release; Miles Brundage’s uptime note shows why multiple independent providers may matter as much as the initial benchmark claim.
    • OpenAI’s computer-use demo shows ChatGPT operating an in-app browser, existing Chrome tabs, and desktop applications; VentureBeat’s enterprise security report supplies the credential warning, while ToolAlignBench tests how tool-using agents behave when safety training conflicts with deployment instructions.
    • CNBC’s Shanghai report details the 29-country cooperation organization and China’s offer of 5,000 training opportunities; Al Jazeera’s account of Xi’s speech records his case for broader access and continued human control.
    • Weco’s Parameter Golf presentation explains how Aiden ran about 1,300 experiments on one H100 and set seven records by testing combinations drawn mostly from public research and other competitors’ work.
    • Techmeme’s Alphabet labor roundup records a 4,500-signature petition seeking severance guarantees and voluntary buyouts, without treating the request as a new layoff announcement.
    • CNBC’s Memphis investigation follows Colossus from a local dispute over noise and emissions into a precedent cited by lawsuits, zoning changes, and policy campaigns elsewhere.
    28 min
  • The Assistant Asked for the Same Door

    Europe is forcing a concrete test of AI competition: can a rival assistant reach the same Android capabilities and search data that make Google's own products useful?

    • The Verge on the EU's Google decisions reports the new Android interoperability and Search-data obligations, where implementation will decide whether comparable access changes the market.
    • The European Commission's Android case summary details access to invocation, ambient sensors, on-device actions, and first-party services that rival assistants currently lack.
    • CNBC on TSMC's second quarter connects a 77.4 percent profit increase with higher spending plans, while the reported US fab expansion remains a pledge rather than finished capacity.
    • Nvidia's Jetson Thor announcement pairs new edge computers with the four-billion-parameter Cosmos 3 Edge model, making the company's robotics offer a hardware-and-software package.
    • The Verge on xAI's lawsuit reports the company's allegations against a Grok user and raises the unsettled role of civil litigation in severe model misuse.
    • The Action Rebinding paper shows how an Android app with no sensitive permissions can redirect a GUI agent between observation and action.
    • The CAVA paper proposes canonical action objects and attestations so approvals can be bound to what an agent executes across different runtimes.
    • Thinking Machines Lab's Inkling release supplies the model architecture, training scale, multimodal inputs, and open weights that were missing from the day's secondary coverage.
    • Axios on frontier-lab regulation compares the different institutions proposed by Demis Hassabis, Sam Altman, and Dario Amodei while a disclosed political contribution makes the policy effort tangible.
    30 min
  • When a Week Away Looked Like a Bad Score

    A lawsuit against Meta asks whether missing work data during protected leave became a negative score, and who is accountable when people still approve the final list.

    • The Guardian’s report on the Meta lawsuit lays out the workers’ allegations, Meta’s denial, and the July 22 deadline that makes the dispute urgent.
    • The Guardian’s report on the publishers’ Google suit follows the alleged path from books supplied for search and sales into commercial Gemini training.
    • Techmeme’s summary of Reuters reporting on xAI’s turbines connects data-center growth to permits and the communities beside the generators.
    • Techmeme’s PJM power-cost item puts a $6.3 billion estimate on the capacity costs attributed to data-center demand.
    • Techmeme’s summary of the data-center power pledge covers the voluntary effort to assign new generation, grid work, and unused reserved capacity to the companies creating the demand.
    • Tailscale’s security bulletin explains how a username beginning with a dash could be parsed as a command option and open a root session.
    • CNBC’s ASML earnings report gives an upstream demand check through guidance and planned lithography capacity.
    • The PalmClaw paper replaces long screen-control sequences with explicit phone tools and reports the resulting task and latency measurements.
    • RunAnywhere’s benchmark-runner announcement adds the useful ability to test a chosen model on the phone that will run it.
    26 min
  • New York Paused the Megawatts

    New York has paused permits for new hyperscale data centers while it decides how their power, water, emissions, and local costs should be governed. The episode follows that one-year bargain, then turns to frontier-model inspection, Nvidia's customer screening, new agent-security methods, GPT-5.6 on Bedrock, and Australia's live copyright dispute.

    • The Verge on New York's data-center moratorium explains the statewide pause for facilities drawing at least 50 megawatts and the standards regulators now have to write.
    • Axios's interview with Demis Hassabis sets out a FINRA-inspired body that would inspect frontier models before release and could eventually delay them.
    • Techmeme's Nvidia customer-screening roundup makes export enforcement concrete through the reported removal of more than half of approved buyers in Singapore, Malaysia, and Japan.
    • Rest of World's Gulf procurement report shows why abundant capital still doesn't produce an independent training stack when Nvidia, CUDA, advanced memory, and foundry capacity remain scarce.
    • The ANCHOR paper proposes persistent, law-grounded misuse audits for long-running command-line agents and reports that direct refusals collapse under adaptive multi-turn pressure.
    • Techmeme's context-bombing roundup describes a defensive use of prompt injection that places refusal-triggering instructions beside cloud secrets.
    • The Mako paper reports fresh-flag verification across 104 web-security challenges while withholding exploit payloads and tool code.
    • AWS's GPT-5.6 Bedrock announcement adds first-party billing, commitment credit, prompt caching, and the Responses API to the models' enterprise purchasing path.
    • The Guardian's interview with Ed Husic captures the Labor dispute over copyright, payment for creative work, and whether AI companies can govern their own training practices.
    28 min
  • Five Gigawatts Need a Town

    Meta's five-gigawatt Louisiana expansion puts a town, a utility, and a public balance sheet inside the AI compute story.

    • CNBC's Hyperion report establishes the more-than-$50-billion expansion, while Meta's announcement describes the five-gigawatt plan, new generation, utility agreement, and local benefits.
    • Phaidra's AI Engineer talk shows why its infrastructure search failed as the equipment tree grew and how deterministic set operations replaced exact model search.
    • Nathan Lambert's open-model warning separates an undefined licensing risk from any policy the White House has formally proposed.
    • The economists' letter report documents a large policy coalition while leaving its requested remedies unresolved.
    • Design Arena's GPT-5.6 Sol result and Ploy's migration report give two narrow measurements: one ranking and one production workload.
    • Rest of World's India report explains why a dispute over WhatsApp usernames has become a test of government influence over encrypted product design.
    28 min
  • Who Gets to Say the Work Is Done?

    AI can produce code, reports, and decisions faster than people can inspect them. Today, Lenar and Damra look at who absorbs that verification cost and what a credible claim of completion now requires.

    • The open-source maintainer report shows how low-effort AI contributions can consume scarce review time, while different projects are responding in very different ways.
    • Paperclip's talk on completion treats done as a bundle of claims about the artifact, evidence, reviewer, authority, remaining risk, and next action.
    • The agent CI/CD talk identifies handoffs as the points where a multi-agent system can report success while passing damaged or incomplete work downstream.
    • The Grok Build wire analysis reports that version 0.2.93 uploaded tracked repository contents and Git history even when the agent was told not to read files.
    • The Guardian's Australia report details the contested proposal to exchange copyright concessions for datacentre investment ahead of Anthony Albanese's policy speech.
    • Indeed Hiring Lab's jobs analysis finds a software-posting rebound concentrated in senior and AI-titled roles, which complicates both collapse and boom stories.
    • Machinecraft's Eira case study describes a factory memory built from quotes, drawings, payment schedules, and email, with people retaining authority to send.
    • The LingBot release summary reports a video-action model, few-shot adaptation, and a large latency reduction for robot control.
    23 min
  • The Talent War Hits the Hardware Lab

    Apple's trade-secret suit against OpenAI turns the AI hardware race into a legal and institutional story: who can hire whom, what knowledge walks out the door, and how labs organize safety, policy, and infrastructure around a much bigger product ambition.

    • The Verge and Axios frame Apple's allegations against OpenAI as a trade-secret case about hardware staff, confidential material, and accelerated device plans.
    • CNBC, plus Techmeme's safety-leadership roundup, put OpenAI's internal reorganization in the same week as the lawsuit: execution, research, and safety are becoming one institutional design problem.
    • CNBC's UAE chip-export report gives the policy segment its hard surface: export control is a collection of approvals, allies, investment ties, and enforcement boundaries.
    • Techmeme's misuse roundup, Forbes on AI ransomware, and CISA-linked reporting on public repo credentials move the safety story from abstract model policy to concrete abuse paths.
    • SK Hynix's market signal, the reported shortage warning, and two AI Engineer talks close the episode on capacity and craft: memory supply on one side, human-readable agent work on the other.
    22 min
  • Work Agents Learn the Office

    OpenAI and Meta both spent Thursday showing agents that can work across files, apps, tools, and long projects. Today's episode follows the product surface first, then the policy and research work trying to catch up to agents that no longer stay inside a single chat window.

    • OpenAI's ChatGPT Work demo showed finance analysis, local files, browser tabs, app control, and shareable hosted sites as one work surface.
    • OpenAI's GPT-5.6 model demo introduced programmatic tool calling, higher reasoning levels, and subagent delegation in Codex.
    • Meta's Muse Spark 1.1 launch post framed its new model around a million-token context window, multi-agent orchestration, computer use, and the public Meta Model API.
    • Techmeme's Financial Times summary reported OpenAI and Google serving advanced models to Singapore-based subsidiaries of Chinese tech companies, a concrete update to this week's model-access debate.
    • The European Commission preliminarily found Facebook and Instagram in breach of the Digital Services Act over addictive design features including infinite scroll, autoplay, notifications, and recommender systems.
    • The harness-engineering paper and the context-graph paper gave the research counterpoint: agents need contracts, source boundaries, monitors, and event-driven context before they can be trusted with enterprise work.
    24 min
  • Coding Models Meet Their Test Bench

    Grok 4.5 gives the day a coding-model lead, but the stronger tension is measurement: the model market is moving faster than the tests, sandboxes, voice interfaces, and power equipment around it.

    • SpaceXAI's Grok 4.5 post anchors the release as a coding-and-agent model rather than a general chatbot update.
    • TryAI's build-off gives the launch a practical counterweight by comparing Grok 4.5, GPT-5.5, and Claude on the same app-building tasks.
    • OpenAI's GPT Live 1 demo shows a full-duplex voice interface with interruptions and delegated reasoning, which makes the interface architecture the story rather than another access update.
    • OpenAI's SWE-Bench Pro note, Databricks' codebase benchmark, and AgentLens point toward coding-agent evaluation that inspects trajectories, not only pass or fail.
    • Techmeme's transformer-lead-time item, Meta's Alberta data-center report, Iluvatar CoreX coverage, and Positron funding coverage keep the compute story grounded in power equipment, geography, and capital.
    • AWS's Claude Apps Gateway, Latent Space's Modal interview, and AI Engineer's agent sandbox talk show the operational layer forming around agents.
    • CNBC's election-spending report and Nathan Calvin's release-authority note keep the policy update separate from yesterday's GPT-5.6 access story.
    27 min
  • The Release Gate Opens

    OpenAI's GPT-5.6 rollout moved from constrained preview to public launch timing, while the day's security and hardware stories showed how much now sits around a model release: authorization, source-code access, chip commitments, and policy scrutiny.

    • OpenAI's GPT-5.6 announcement sets Thursday as the public launch for Sol, Terra, and Luna, with preview access expanding globally now.
    • Axios and CNBC report the rollout followed U.S. Commerce Department clearance after earlier government-requested limits.
    • Noma Security's GitLost report puts agent security in concrete repository terms: an AI coding agent could be steered toward private-code exposure.
    • The MCP concealment paper shows how hidden Unicode TAG-block payloads can vanish from a human approval view while still reaching the model.
    • The off-host authorization paper argues for moving authorization out of the agent host, with per-message identity, argument-level policy, and hash-chained audit records.
    • Tim Cook's Broadcom post and CNBC's report turn chip supply into a signed U.S. manufacturing commitment rather than a vague infrastructure promise.
    20 min