Click any tag below to further narrow down your results
Links
Nous Research's open-source Hermes Agent is now available inside Microsoft Teams, the Microsoft Store, and Azure, aimed at companies that want a self-improving agent for every employee. The pitch is that successful workflows get saved as reusable skills the whole team can use.
- Hermes Enterprise runs in Microsoft Teams, Hermes Desktop for Windows (via the Microsoft Store), and Azure Container Apps Sandboxes, where agent tasks run in isolated microVMs that start in under a second.
- Hermes turns a successful approach to recurring work into a reusable skill, and a Collective Wisdom feature lets employees review and publish those skills for their team.
- Companies with strict data residency or procurement needs can run Hermes in their own Azure tenant.
OpenAI is pitching agents that operate EHRs and payer portals by clicking and typing, so they can handle tasks like MRI authorization that APIs don't reach. The post shares evaluation results on helpfulness, safety, and cost, plus early deployments with Cigna, CVS Health, Abridge, and Ambience.
- A demo agent takes a patient's message and visit notes, pulls prior imaging and PT records from the EHR, and drafts an MRI order, encounter note, and patient reply for clinician review.
- CVS Health reports an agent cleared a months-long document-sending backlog over a weekend and handles system steps for tens of thousands of consults a day, with validation checks and human oversight.
- Cigna case managers previously had to navigate more than 10 separate systems to gather member context; computer use now pulls that together in seconds.
Google's Gemini Notebook can now turn your notes and sources into interactive sites, just by typing "make me a site" with a description of what you want. It's rolling out on the web to users 18 and up, including Workspace subscribers, with mobile coming later.
- Typing "make me a site" in a Gemini Notebook chat generates an interactive site from your sources, with examples like a periodic table for exam prep or a quarterly sales dashboard.
- Finished sites open in a new tab or can be shared with others.
- Available globally on the web now for users 18+ and Google Workspace subscribers; the mobile app is coming soon.
Perplexity is recruiting for Frontier Research, an initiative to build software that runs the full research loop (asking questions, gathering evidence, forming hypotheses, testing them) and make it available to anyone. The work spans multi-agent orchestration, pre-training, continual learning, reinforcement learning, systems, and new evaluations that measure discovery rather than recall.
- The core bet is that automating scientific research is the next big advance, with agents that search widely, go deep, propose experiments, and critique each other.
- Perplexity says it will back the effort with more than 10,000 GPUs and publish papers, models, data, and benchmarks, with results shipping into a product used by millions.
- It argues current evals reward proxy skills rather than general ability, so it plans new benchmarks that test qualitative understanding and discovery.
The author argues that AI has pushed formal verification into the mainstream, but "formally verified" is a misleading label because proofs only cover what's specified, and the trusted base (definitions, models, provers) is still hard to check. He's optimistic about proving specific properties of software, especially single-threaded code, and thinks random testing will help validate the models used.
- Formal verification is now economically valuable and widely discussed, with AI generating large volumes of proofs, but that volume makes it more important to scrutinize what is actually trusted.
- Testing produces reviewable bug reports, while proofs rely on hard-to-verify components like specifications, machine models, and provers, so "formally verified" should always come with caveats about what's proven and what's assumed.
- Software with user-facing, "bouba" layers resists uniform correctness proofs, but specific properties of single-threaded programs are likely provable within years, with random testing helping to validate the models.
The author, a former theoretical physicist, argues that superintelligence won't cause the explosive economic change rationalists predict, because physical, regulatory, and informational bottlenecks cap how much faster the overall world can move. He uses Amdahl's Law and a widget-factory example to show that speeding up one step rarely speeds up the whole process by much.
- Amdahl's Law: if a process has several sequential bottlenecks, infinitely speeding up one still leaves a constant-factor overall speedup, so exponential gains require repeatedly removing the current bottleneck.
- In the factory example, even with omniscient AI agents working 24/7, a month-long fix only drops to about a week because of shipping times, lab processes, and installing physical hardware.
- Some bits of information are very expensive to acquire (like long-term side-effect data), and regulation and political will are slow speed limits that intelligence doesn't remove.
Bloomberg reports that Anthropic's initial public offering has drawn several times more orders than shares on offer, driven by large anchor-size indications of interest. If accurate, it would be among the biggest first-time share sales on record.
- Orders for Anthropic's IPO are reportedly multiple times the number of shares available.
- Demand was led by anchor-size indications of interest, meaning large institutional investors committing big amounts early.
- Sources say the deal could be the largest first-time share sale ever, though the information is not yet public.
OpenAI's revenue passed $10 billion in the three months ending in September, according to people familiar with the company's investor call. The growth was driven by demand for its latest AI software.
- Quarterly revenue reached $10 billion for the first time, up nearly 50% from the prior quarter.
- Revenue almost tripled compared with the same quarter a year earlier.
- Greg Brockman, Sarah Friar, and Thibault Sottiaux discussed the figures on a Tuesday investor call. The numbers are not public, per the sources.
@lennox_saint on X
A user on X says you can get better results from Claude Code by picking the Fable 5.2 model with /model and setting sub-agents to Opus 5.5, so they don't copy the main model's output. The post is a short how-to with a prompt template asking for a title, summary, key points, and tags.
- Steps: open Claude Code, run /model and choose Fable 5.2, then set sub-agents to Opus 5.5.
- The poster says sub-agents set to the same model copy the main model's output, which is why they recommend the split.
- The post offers a structured prompt template for big jobs, asking for a highlight, summary, key points, and tags.
The Electronic Frontier Foundation filed a FOIA lawsuit to get records on WISeR, a Medicare program launched in January 2026 that uses AI to review prior authorization requests. The released documents include vendor contracts, accuracy and bias testing requirements, and status reports on how the program is running.
- WISeR uses AI to assess prior authorization requests from Medicare beneficiaries, and experts and lawmakers have warned it could cause serious patient harm, including wrongful denials of care.
- Little is public about how the algorithms work or what data they were trained on, and it is unclear whether safeguards exist against algorithmic bias or privacy violations.
- The released records include participation agreements with six tech companies, reviewer criteria, a data reporting guide on vendor audits, and an actuary estimate of the program's impact.
Chris Neumann argues that founders of startups outside the Bay Area can't grasp how fast AI is reshaping how companies build products by reading about it, so they need to visit Silicon Valley in person. He says many of the founders he talked to were reluctant to make the trip even as their teams struggled to keep pace.
- A Series B founder reported that AI changed how his team builds product so quickly that some early employees couldn't adjust, even within weeks.
- Neumann says the founders he speaks with ask themselves "how do I keep up?", and his answer every time is to come to SF and see the changes firsthand.
- Many founders he advised said they would "look into it next year," while those who booked flights said afterward they had no idea how fast things were moving.
Google is rolling out Gemini 4 Carbon, its most capable model, aimed at hard tasks like multi-step coding, scientific research, and cybersecurity work. It costs more to run than the lighter Gemini 4 Argon, but Google says it's cheaper per task than competing top models on several benchmarks.
- Priced at $2 per million input tokens and $10 per million output tokens, the same introductory rate as Argon, and Google claims more than 2x lower cost per task than the next best external model on CursorBench 4.0, WANDR, and AutomationBench.
- Google previews /swarm in Google Antigravity, which coordinates multiple agents with Carbon handling architecture and auditing while lighter models do parallel work.
- In cybersecurity, Google reports Carbon found 0-day vulnerabilities in two major SaaS providers in under two hours, though these claims come from Google and its partners rather than independent testing.
Thomas Wolf, a startup founder, argues that AI's strong benchmark scores don't match its limited economic impact because real work depends on judgment and agency, which benchmarks rarely measure. He ties this to a drop in entry-level hiring, since execution is the part of junior work AI handles best, and that junior work is where people usually build the judgment that later makes them senior.
- A Stanford analysis found employment for workers aged 22–25 in the most AI-exposed occupations fell about 6% between late 2022 and mid-2025, while older workers in those jobs gained 6–9%.
- Models now solve around 75% of real-world coding tasks on SWE-bench, yet the Remote Labor Index shows the strongest agents succeed on only about 2.5% of paid freelance projects.
- Wolf's hiring framework splits workers into execution, judgment, and agency; AI is advancing fastest on execution, so the entry layer gets squeezed first, weakening the pipeline for future senior people.
Apple says some Apple Intelligence features run on server-side models, so heavy use will face limits that reset after a timeout, with higher access available for a fee starting with the 2027 releases. The article also caps how many cameras get Home app video summaries based on your iCloud+ plan tier.
- Starting with 2027 software, features like Siri AI, Image Playground, and Clean Up/Extend photo tools get enhanced server-side access, but usage limits will apply later.
- Limits vary by feature, request complexity, system demand, and policy, and exceeding one pauses that feature until a timeout ends.
- Home video summaries are capped at 1 camera on the 2 TB plan, 2 on the 6 TB plan, and 5 on the 12 TB plan.
The author argues that making software cheap is likely to mean much more software, as Jevons predicted for coal in 1865, but that whether the people who build it gain depends on what the new tool does to their specific job. Historical cases like farming, light, bank tellers and telephone operators show the paradox holds for output but can fail for workers.
- Jevons argued in 1865 that efficient engines would increase coal use, and that cheaper products would widen employment, which the coal boom bore out for about fifty years before output collapsed.
- The price elasticity of demand decides the outcome: food (elasticity roughly 0.27 to 0.81) saw the savings leave the industry, while light behaved as if elastic and use rose thousands of times over.
- Bank tellers survived ATMs because branches multiplied, but the job shrank from 600,500 in 2008 to 339,200 in 2025 once mobile banking removed the need to visit, and telephone operators lost 50 to 80% of their jobs almost immediately after mechanical switching.
- Software output already looks like coal in 1865: Google's token processing rose about 330 times in two years, and new repositories and app submissions climbed sharply.
The author argues that AI makes cheap to fake the signals we use to judge quality, such as GitHub stars, polished writing, and homework passing autograders, so good work stops getting rewarded and a market for lemons forms. He uses signaling theory to explain why this change is so fast and disorienting compared with earlier, slower erosion.
- Shortcuts like stars, clean READMEs, and passing tests worked because faking them used to cost nearly as much as doing the real work; AI removes that cost gap.
- Buying GitHub stars already cost about $0.07–0.20 each, and AI now makes fake accounts with realistic histories easy to produce, hiding the fakery better.
- When signals collapse, evaluation gets more expensive, and buyers, employers, and reviewers stop paying a premium for good work, so poor work dominates, as in Akerlof's 1970 used-car lemons paper.
- Past signal erosion (biological mimicry, CAPTCHAs, pass/fail grading) was usually slow, but AI is breaking many signals across open source, education, and research at once.
This is Apple's developer landing page for macOS, pitching the tools developers can use to build apps that plug into Apple Intelligence, Siri, and the system. The main pieces are App Intents for surfacing app content to Siri, the Foundation Models framework for on-device and cloud AI features, and graphics and video APIs for games and editing apps.
- App Intents lets apps expose content and actions to Siri and Spotlight through schemas, so users can act on them with natural language and no custom phrases to define.
- The Foundation Models framework is a Swift API giving access to Apple's on-device and Private Cloud Compute models, and it can also run any provider that conforms to the Language Model protocol.
- The Video Effects API offers machine-learning tools like Motion Blur, Frame Rate Conversion, and up to 4x super-resolution upscaling for video editing apps.
A long NYT profile of Anthropic CEO Dario Amodei, 43, written as Anthropic prepares one of the largest public listings of the year. It tries to explain the ideas shaping how he thinks about AI, drawing on his background and worldview.
- The piece ties Amodei's personal intellectual formation to the strategic choices behind Anthropic.
- It arrives just as Anthropic is set to go public this month, which gives the profile its timing.
- It is framed as an attempt to show how the person at the center of the AI race actually reasons about the technology.
Conductor, the agent workspace tool, is being acquired by Meta, so its cloud service goes read-only on November 13 and shuts down on November 20. The local Mac app keeps getting model updates and bug fixes until early 2027, and this release also adds custom completion sounds and several UI fixes.
- Conductor Cloud goes read-only on November 13 and shuts down on November 20, so cloud users need to move their work soon.
- You can transfer a cloud workspace and its chats to your Mac by right-clicking it and choosing Transfer to local, or export all data from Settings → Account.
- The local Conductor app gets new models and critical fixes until early 2027.
@ClaudeDevs on X
Claude Code's background subagents used to re-read their entire conversation every time they updated their status line, which ate into usage limits. Anthropic now has Haiku 5.5 write those updates from short notes, cutting the cost of each update by about 99%.
- Background subagents post a status update every 30 seconds.
- Previously each update re-read the subagent's whole conversation, which consumed a lot of usage.
- Haiku 5.5 now writes the updates from short notes, making each one roughly 99% cheaper.
Randy Au argues that LLMs can take over tedious coding work in data analysis, but exploratory data analysis still needs a human who understands the data's context and intent. He says current models make poor analytical decisions and that the act of poking at data is how analysts sharpen their hypotheses.
- Exploratory analysis depends on intent: the same numbers look "correct" or "weird" depending on what you plan to do, and an LLM has no access to that purpose unless you spell out a huge amount of business context.
- Coding agents lack a feedback signal for analytical quality, so they confidently run off on flawed approaches. Au says he had to micromanage them to the minute detail, and the promised analysis often didn't match what was delivered.
- Hypothesis generation happens through hands-on iteration with the data. Offloading that work skips the process where questions get refined and interesting correlations get noticed.
OpenAI is folding the Chat and Work tabs in the ChatGPT desktop app into one screen, so you can go from idea to finished output without switching modes. Codex stays separate, and chat and Codex will draw from a shared token pool.
- The Chat and Work tabs are being merged, starting with a limited group of users for feedback before a wider rollout.
- You can bring in files and plugins and have ChatGPT build a website, make a presentation, or act on the web while you review its work.
- ChatGPT and Codex will share one token allowance, which OpenAI says won't affect most users, though heavy users may hit limits and wait for resets more often.
Elliot Morris argues that prototypes used to force engineers to communicate clearly, surface real problems, prove competence, and engage with reality, but LLM-built prototypes break most of those benefits. He thinks the useful future prototype is a sharp, single-thesis experiment rather than a big, sprawling demo.
- Prototypes once delivered four bundled benefits: clear communication, early problem discovery, a visible competence signal, and forced contact with reality. LLM prototypes keep only the first of these reliably.
- LLM output tends to drift toward a generic "average" answer, so the builder can end up communicating something they didn't intend without noticing.
- Experienced engineers make rougher, smaller prototypes because they test one narrow thesis. Morris suggests the key question is now "what specifically does this prove, and why should I care?"
@claudeai on X
Anthropic's Claude app on iPad and iPhone now opens anything Claude creates beside the chat instead of on top of it. You can expand it to full screen when you want a closer look.
- Created content opens side by side with the chat on iPad and iPhone Duo, rather than covering it.
- Users can switch to full screen for a closer view of what Claude made.
- The change is one of several upgrades to the Claude mobile apps shipped this week.
Brands are using AI to create digital clones and face swaps of models, often under broad contract language that gives them rights to reuse and train on those images without extra pay. Reuters found the trend hitting ecommerce modeling work hardest, with New York's 2025 law the main legal barrier so far.
- Zalando closed both its Berlin photo studios and cut more than 340 jobs, and About You's Scayle Studios claims a 90%+ production-cost reduction using AI-generated models.
- Unsigned contracts from Diesel, Calvin Klein, and L'Oreal sought rights to swap faces, create digital clones without added pay, and use model images to train AI tools.
- Since June 2025, New York's Fashion Workers Act has made it illegal to create or modify model replicas without explicit consent, while most other regions have no binding rules.
ElevenLabs has released Cue, a forced alignment model that maps every character in an audio file to a precise timestamp. The article's case is that it beats open-source aligners on long files, noisy audio, language coverage, and speed.
- On full recordings, Cue averages 179 ms of timing error, while the open-source tools tested ranged from 283 ms to over 6,000 ms.
- It handles noisy English recordings with a 57 ms average error, versus 65 ms to 920 ms for the others, and processes an hour of audio in about half a second.
- One model covers 29 benchmark languages in a single pass, so long files don't need chunking, which is where timing errors usually build up.
This is a product page for an AI tool that generates 360° skybox panoramas from a text prompt and up to three reference photos. You can preview the result on a sphere, adjust yaw, exposure, and seams, then export a PNG panorama or six cubemap faces.
- Output is a 2:1 equirectangular image in 1K, 2K, or 4K, in Realistic, Cinematic, or Illustration styles.
- Exports are a PNG panorama or a ZIP of six cubemap faces, with a separate Cubemap Converter for other face sizes and naming.
- It generates visual environment images, not measured HDR lighting, so captured HDRIs are needed where real light intensity matters.
Chris Neumann, a venture investor, says this fall's fundraising market has shifted hard toward founders pitching bold, credible companies, especially in AI and deep tech. Incremental B2B businesses are getting little attention, and he advises most founders not yet fundraising to wait until the new year.
- Founders have adapted to VC demand for "unapologetic ambition," and the quality of companies raising this fall is unusually high.
- Neumann met five repeat-founder teams raising or preparing $20-30M Seed rounds in one week, and every one was credible.
- Warm intros are taking 2-3 weeks to turn into first meetings, so companies starting now risk missing the window before year-end, and he suggests focusing on customers until Q1.
Anthropic has filed a registration statement with the SEC for an initial public offering of its Class A common stock. The shares would trade on Nasdaq under "ANT," but the share count and price range aren't set yet.
- Anthropic, PBC filed a Form S-1 with the SEC for a proposed IPO of Class A common stock.
- The planned listing is on the Nasdaq Global Select Market under the symbol "ANT."
- Goldman Sachs and Morgan Stanley are the underwriters handling the offering, and the terms are still undetermined.
Venkatesh Rao uses Joyce Appleby's history of capitalism to argue that fears about AI replacing human thinking repeat the zero-sum logic of pre-capitalist mercantilism. He suggests that just as agricultural gains freed most people from farming, AI could free people from cognitive labor and let them become consumers of intelligence rather than its producers.
- Mercantilism assumed wealth was fixed and that ordinary people couldn't be trusted with freedom; Rao maps both assumptions onto "cognitive mercantilism," where intelligence is hoarded and treated as scarce.
- Appleby documents a "industrious revolution" in Britain, with work days rising from about 250 to 300 a year, driven partly by voluntary desire for new consumer goods like tea, sugar, and tobacco, well before the Industrial Revolution.
- Rao argues the analogous path for AI is that most people won't need to produce intelligence to benefit from it, just as few people today farm yet eat abundantly.
Electric Plant, a startup, has built a small sensor that picks up the electrical signals plants emit and uses AI to interpret them as data on hydration and stress. The company's CEO, Justin Mast, describes the device as a kind of dongle.
- Plants produce electricity, and their electrical signals shift in response to environmental stimuli.
- The sensor decodes those signals to gauge conditions like hydration and stress.
- The device is pitched as a simple add-on rather than a complex monitoring system.
@techNmak on X
Graphify is an open-source GitHub project that parses a codebase once into a persistent knowledge graph, so AI coding agents can trace how functions, files and services connect instead of repeatedly grepping and re-reading source. It plugs into assistants like Claude Code, Codex and Cursor through hooks, instructions or MCP.
- It uses tree-sitter to parse about 40 languages locally, extracting functions, classes, imports, calls and cross-file relationships without embeddings, a vector database or LLM calls.
- Every relationship is tagged EXTRACTED, INFERRED or AMBIGUOUS, and community detection surfaces related subsystems and central nodes.
- The graph can be committed to Git, updated incrementally, and checked against pull requests to flag changes that touch related areas.
A mother who works at fast-growing AI company Lovable argues that 996 schedules (9am to 9pm, six days a week) are a bad strategy, both for the people working them and for companies that filter out parents and others with real lives outside work. She says boundaries only hold if companies build cultures that protect them and if people stop overriding their own limits.
- Working 996 gets treated as a badge of honor in AI companies, but she argues it confuses long hours with ambition and talent.
- Companies that require 996 as the default borrow against employees' future health and relationships, and that cost never shows up on the balance sheet.
- Excluding parents and caregivers narrows the range of perspectives building the product, and constraints outside work often make people sharper at prioritizing.
The author fine-tuned a small 360M-parameter language model to simulate a terminal with no real shell behind it, predicting how the screen changes after each keypress. It can create files, run vim and tmux, and play snake, though it only sometimes remembers earlier commands.
- Instead of generating text token by token, the model predicts screen updates, with a COPY operation for scrolling that cuts a full redraw from roughly 3,000 tokens to about three tokens per row.
- Training on its own mistakes (a DAgger-style repair approach) raised PresetBench from 24/47 to 38/47 at int4.
- Decoding rows in parallel brought full-screen redraws from up to 12 seconds down to something usable, making near-realtime performance possible on consumer hardware.
- Memory is still unreliable: the model read back correct values in only 9 of 13 file tests, and combining commands in new ways often breaks.
This is a web tool that uses AI to swap out words in an existing image, like a price on a poster or a name on an invitation, while trying to keep the original font style, color and background. It's aimed at people who have a finished graphic but have lost the source design file.
- You select the text in an uploaded JPG, PNG or WebP (up to 10 MB), describe the replacement, and get a separate edited copy; up to eight edits fit in one 40-credit task.
- The AI matches original lettering, size, color and background, but an exact font match isn't guaranteed, so logos, decorative type and tight spaces need close checking.
- New accounts get 80 free credits (two edits), and it works on flat images rather than producing editable layered text, so it's a patch job rather than a replacement for the original design project.
The Register reports that the creator of OpenClaw says he spent about $10 million on AI tokens in a single month. The piece is one of several items in a roundup of AI, security, and open source news.
- The OpenClaw creator says he spent roughly $10 million on tokens in one month.
- The figure points to how expensive heavy agent-style use of AI models can get.
- The article itself is a headline roundup, so the OpenClaw item is only one brief entry among many on AI, infosec, and FOSS.
Andreessen Horowitz is expanding its crypto fund into a broader "frontier" fund that backs long-horizon technologies, from AI and new hardware to blockchain-based payments for machine networks. The firm says the crypto label matters less now that blockchains sit mostly in the background.
- The fund is being renamed from a16z crypto to a16z Frontier, but crypto investing and crypto regulatory advocacy continue.
- The firm argues companies should be defined by what they do, not by the technology they use, since Wall Street and fintech now routinely build on blockchains.
- Its thesis is that AI will enable "native" use cases and machine-to-machine networks, and that stablecoins and programmable blockchains can give those networks economic capabilities.
This is a landing page for an online car customizer that lets you mock up body kit styles on your own vehicle before you buy anything. The pitch is that a visual preview gives you something concrete to talk through with an installer.
- You can compare subtle aero, widebody, and track-style body kits on your own car before committing.
- The preview is meant as a discussion aid for installers, not a finished design or quote.
This is a browser-based tool suite that turns text, photos, video and hand-drawn scenes into character art. Basic conversions run on your device and export as plain text, PNG, SVG or GIF without an account.
- Text banners use FIGlet fonts, and photos are converted by sampling brightness into a character ramp, so the two produce different results.
- Core workspaces (image, text, drawing, animation, pixel art, photo mosaic) are free, and basic exports need no credits or sign-in.
- Premium effects and AI rendering require an account, and AI rendering uploads a copy of your artwork to an external provider.
- Projects can be saved as JSON files for portability, since browser draft storage stays on the original device.
This is the landing page for Roblox Shirt Maker, a browser tool for creating Roblox shirts, pants, and graphic T-shirts from templates or AI-generated designs. It lets you edit panels, preview on an R6 avatar, save drafts, and download PNGs for upload to Roblox.
- Three clothing types are supported: classic shirts and pants export 585 × 559 PNGs, while graphic T-shirts export transparent 512 × 512 PNGs.
- Editing, previewing, and saving projects are free; AI generation needs an account and credits, with 180 starter credits (three generations at 60 each) for new accounts.
- Drafts are stored in the browser, so you need to export a project file to move edits to another device or keep a backup.
The author argues that frontier AI models are getting good at generating CAD and designing parts, but physical hardware is hard to automate because the link between digital designs and the real world is lossy in both directions. Writing a design into atoms and reading measurements back out both introduce errors that models will have to work through.
- Hardware work breaks into levels (101 to 501), from sourcing COTS parts to building a million units, and each level needs different skills that have to be solved separately.
- Design for manufacturing is the hard part: models may produce ideal parts, but making them buildable and cheap requires constant real-world interaction.
- Real-world data is non-uniform, sparse, expensive to collect, and slow, so measuring what is actually happening on a factory floor or supply chain is difficult.
Chris Kiehl pushes back on the gloom around Hacker News, arguing that while AI has changed the economics of paid programming, coding for its own sake is still worth doing. He describes his weekend projects, a buggy Breakout clone, Java type-safety experiments, and a frustrating dive into recursion schemes, to show that learning hard things is rewarding without any business payoff.
- Programming for work and programming for enjoyment are different things; AI changing the value of hand-written code at a job doesn't remove the reason to do it for fun.
- Kiehl admits his projects have no clear payoff: a bad Breakout clone, Agda-inspired Java experiments, and recursion schemes that left him unsure what they were even good for.
- The "there is no point" feeling comes from constant exposure to doom-heavy online discussion, and it fades once you log off.
Anthropic is offering free, fully model-generated security scans to eligible open source projects through a new program called OSS Scanner, with reports delivered without human triage. Early testers say the findings were mostly valid, though the company admits some reports will be wrong.
- Anthropic's models found over 29,000 candidate vulnerabilities in six months, but only about 6,000 have been manually reviewed, so human review is the bottleneck.
- OSS Scanner reports go out unverified, but in a pilot, 85 of 97 critical and high-severity findings (88%) met the bar for Anthropic's disclosure process, and only one was a false positive.
- Eligible projects get reports with a reproducer, an explanation, and often a candidate patch; several maintainers said the reports were high quality, and some chained bugs into unauthenticated remote code execution.
Men who work in tech, especially at firms like Palantir and Tesla, say their jobs now trigger suspicion or rejection from potential partners because the industry is seen as politically and morally compromised. The piece follows several daters and a matchmaker to show how workers hide or justify their employers to get dates.
- Daters describe brutal reactions to naming a Palantir or Tesla employer, including one woman calling a man "the devil" and friends asking him not to mention the job on nights out.
- Men are hiding or fudging their jobs: one Palantir engineer says he lists his work as "AI" and invents explanations, while a Tesla employee found he got more matches on Raya by leaving his employer off his profile.
- Matchmaker Amy Laurent now pre-empts objections when pitching tech clients, saying "Listen, I have a great guy… He's in tech," and women cite the industry's record on misinformation, addiction, labor, and wealth inequality as reasons to refuse.
The New York Times profiles Terence Tao, the mathematician who became a tenured professor at 24, and why he has ended up as the field's unofficial spokesman on AI. The piece argues that established academics like him feel less threatened by AI than younger researchers just starting out.
- Tao says senior academics are not as threatened by AI as people beginning their careers.
- His celebrity has made him the de facto public voice for mathematics on AI, and that role now takes up most of his time.
- The math community is looking to him for answers, which the headline frames as a "war footing."
Anthropic is setting aside a share of its stock for a new entity called the Anthropic Public Benefit Fund, with the stake growing as the company's valuation rises. The fund is meant to pursue public-good work in ways the company's internal programs can't, such as large donations or direct payments to groups affected by AI.
- Anthropic commits 5% of its shares now, subject to stockholder approval, plus an extra 2% at board discretion for every 50% rise in company value above a set baseline.
- Combined with the co-founders' 80% equity pledge and an employee equity-match program that about half of staff use, roughly 20% of shares will go to the public good.
- The fund is a wholly-owned LLC rather than a 501(c)(3), and its equity can't be used for lobbying or political giving in the US or abroad.
Google now pays more than 400 publications to license their articles for experimental AI features in Google News, including AI story previews and custom audio briefings. The deal terms are undisclosed, and the pilot pays a flat licensing fee rather than per use in AI answers.
- The pilot has grown to over 400 publications, including CNN, the Financial Times, Axios and Forbes, plus US local papers like the Texas Tribune and Post and Courier.
- Licensed content feeds AI-generated story previews on Google News publisher pages, which carry a note saying the copy "may include mistakes."
- Google says the pilot does not pay based on how much content feeds an AI answer; that per-use model belongs to a separate programme covering AI Overviews, AI Mode and Gemini.
A GNOME developer used Claude with a custom harness to build Linux drivers for unsupported hardware by having the AI analyze USB traffic and Windows driver behavior, successfully fixing a film scanner and laptop camera without access to documentation or original source code.
- Built a working SANE backend for the Plustek OpticFilm 8200i scanner (GL128 chipset) that's faster than the upstream driver added months later, plus a GNOME app called Negative Vibes for scanning film with infrared dust removal and non-destructive editing.
- Created hw-harness, a generic tool that lets Claude compare reference captures from Windows drivers against Linux attempts, turning vague "it doesn't work" into concrete lists of register and control differences to fix.
- Fixed critical bugs in a ThinkPad X1 Carbon's MIPI camera (Sony IMX471) by having Claude identify and patch kernel driver issues and libcamera software ISP problems, raising frame brightness from 31.1 to 90.9 in low-light tests.
TypeSafe released Jev 1.14.0 with a split approach: stable improvements for production use and experimental "danger zone" features (image support, bounding box type, longer context) for developers willing to accept less reliability. The company is also sharing internal intelligence metrics while acknowledging their limitations.
- Image support is now composable across Jev's JSON interface (state, instructions, criteria), though multimodal performance is weaker than text-only and requires careful prompt decomposition.
- New "box" type returns bounding boxes without type errors, representing the first Jev type designed beyond binary decisions and signaling exploration into broader System One Model capabilities.
- Jev supports longer context lengths than previously enabled in production; developers can opt in via a danger zone flag, but TypeSafe intentionally limits context by default because reliability degrades at scale.
- Internal benchmarking data is being shared for the first time, limited to open-weight models only, with explicit caveats that the methodology will evolve and developers should run their own evals.
This is a guide to writing authentic digital love letters, with advice on how to start (memories, gratitude, admiration) and whether to use AI drafts as a starting point. The core message is that the letter should sound like you, not like a template or AI output.
- You can begin a love letter with a specific memory, expression of gratitude, or something you genuinely admire about the person
- If you use AI to draft a letter, you need to revise it until it actually sounds like your own voice
- The best approach is writing every word yourself rather than relying on generated text
AI agents that can manage complex customer journeys over weeks or months—not just single interactions—will let vendors get paid for revenue growth and better outcomes, not just cost savings. This shifts the economics for AI companies and threatens traditional outsourcing models like Red Ventures.
- Long-running agents can now execute multi-step tasks across extended timeframes, handling entire "relay races" like personal injury case management from intake to settlement, eliminating handoff failures and information loss that plague human-driven processes.
- Vendors can move from cost-reduction pricing to outcome-based contracts when they own the full customer journey and can clearly measure results—a plaintiff law firm's settlement amount, an insurer's closed policy, a mortgage lender's funded loan.
- Red Ventures' business model is structurally vulnerable as AI agents replace both the top-of-funnel discovery (personal recommendation agents) and the downstream linear sales funnels that humans traditionally managed, explaining its EBITDA decline from ~$370M to ~$160M.
This is a demo of Fish Audio's text-to-speech tool using a scripted monologue in Uncle Roger's voice — a character known for critiquing improper cooking techniques. The page lets you generate audio from custom text and adjust playback speed and volume.
- Fish Audio offers a text-to-speech platform with character voice models, demonstrated here with the "Uncle Roger" persona
- The tool includes basic audio controls: volume adjustment, playback speed (1x shown), and character selection via model ID
- Users can input custom text (up to 500 characters) and generate dramatic audio with emotional direction tags like [horrified] and <slow> for pacing
Google released Gemini 4 Argon, a new frontier model designed to handle complex multi-step business tasks across software development, finance, legal work, and security. The model can maintain context over long documents, deliver precise results with verifiable citations, and execute autonomous work from a single prompt.
- Argon handles multi-step tasks that previously took hours or days, processing long documents with mixed text, tables, charts, and images into structured outputs with honest reporting of limitations.
- In regulated domains, Argon leads on benchmarks for precision: it tops the Vals Index for financial and legal work, scores 68% on CWE-Bench for security vulnerability remediation, and achieved 20% better task success than earlier models in ServiceNow's testing.
- Google's internal teams used Argon to cut code session prep time by 40%, reduce rollbacks by nearly a third, and compress months of architectural refactoring work into weeks on Chrome's 40-million-line codebase.
- Argon resists indirect prompt injection attacks at 0.7% success rate across 15 adaptive attempts and includes defensive security capabilities while blocking offensive tasks like pentesting.
Judy Faulkner built Epic into the largest EHR vendor by keeping it private and refusing acquisitions, while Larry Ellison's Oracle is pursuing growth through buying Cerner and positioning EHRs as a starting point for automating the entire healthcare ecosystem. Hospitals have sided heavily with Epic's approach so far.
- Epic's ownership structure—Faulkner holds all voting shares and has locked the company into a trust that permanently prohibits going public, acquiring competitors, or being acquired itself.
- Oracle's strategy treats the EHR as a foundation for broader automation across payers, regulators, and public health systems, a fundamentally different market vision from Epic's integrated-system focus.
- Both companies claim AI will reduce clinician documentation burden, but Epic's Agent Factory already has 120 out-of-the-box AI features deployed at 300+ organizations, suggesting different implementation speeds.
Coinbase built an LLM-powered fraud detection agent for its Onramp crypto payment service that sits downstream of existing ML models and rules, reviewing transaction sequences for patterns that emerge across multiple purchases. In live testing, the agent reduced fraudulent transactions by 30% and fraud value by 22%, with projected fraud prevention savings roughly three times the inference cost.
- The agent catches fraud patterns invisible in single transactions by reviewing structured context, historical summaries, and transaction sequences together—something the existing ML model and rules engine miss.
- Selective routing of transactions to the LLM balances fraud detection against latency and cost, with the agent only overriding decisions to block (never to allow) what the upstream checks flagged.
- Newer model versions don't automatically improve domain-specific performance: Part 2 found that newer Opus, Sonnet, and GPT releases all scored lower on recall and F1 for fraud detection, motivating Coinbase to post-train an open-weight model instead.
Mark Zuckerberg and Alexandr Wang discuss Meta's new Muse Wonder AI agent, detailing how they trained it for multimodal perception and agentic capabilities while committing $100B+ to frontier model development without guaranteed success. The conversation covers real use cases—Zuckerberg's MMA coaching agent and Wang's tea-sourcing agent—and Meta's bet that personal AI agents will become the dominant product form factor.
- Meta built Muse Wonder specifically for personal agent use cases, prioritizing multimodal capabilities (video, audio, vision) over pure text/coding, which diverged from industry consensus that coding was all that mattered
- Zuckerberg uses Muse as a real-time MMA coach that processes gym camera feeds and generates tactical feedback within 60-second breaks between rounds, demonstrating the model's video processing and domain knowledge
- Meta committed $100-200B in CapEx before validating whether the approach would work, betting that superintelligence and agentic commerce will drive future product categories beyond smartphones
The attention mechanism, invented in 2014 to solve machine translation, became the core mathematical operation powering modern AI systems like ChatGPT. The mechanism lets neural networks focus on relevant parts of input data, enabling them to handle complex language tasks and scale to larger problems.
- Dzmitry Bahdanau's attention mechanism solved a critical bottleneck in neural machine translation: previous models performed worse on longer sentences because they tried to compress all information into fixed memory before decoding. Attention let networks selectively focus on relevant words, making translation scalable.
- Google's 2017 transformer architecture stripped away everything except attention, proving in "Attention Is All You Need" that this single operation was sufficient for state-of-the-art translation with less computation than prior models. The paper has been cited over 250,000 times.
- Large language models work by predicting the next word across massive datasets; machine translation is essentially a subtask of this general prediction problem, which is why attention mechanisms transfer directly from translation to chatbots, reasoning, and other AI capabilities.
This is Anthropic's hands-on course teaching you how to write better prompts for Claude, with 9 chapters covering prompt structure, common failure modes, and practical techniques. You work through exercises and test examples in a live playground to see how prompt changes affect Claude's responses.
- The course teaches 80/20 techniques to fix common prompt failures and covers Claude's specific strengths and weaknesses
- It's structured as 9 chapters plus an appendix, designed to be worked through in order with practice exercises and an answer key
- The tutorial uses Claude 3 Haiku (the smallest and cheapest model) but notes that Sonnet and Opus are available as more capable alternatives
This is a sign-up page for Zhipu AI's GLM-Legal API, inviting law firms and legal teams to apply for access and discuss how the tool can integrate into their practice. The pitch covers using it for daily legal work, business operations, and client services.
- GLM-Legal is positioned as a tool for routine legal tasks and evolving legal challenges
- The offering targets three use cases: legal practice, business operations, and client services
- Access requires form submission and direct conversation with the Zhipu team about collaboration terms
Mark Seemann argues that fully LLM-driven software engineering could work—producing code as reliable as human-written systems—but legal systems will struggle to assign blame when things go wrong, likely scapegoating whoever signed off on the release rather than the AI that created the defect.
- Code review and human inspection were never foolproof either; the real question isn't whether LLM code is perfect, but whether it's better than human-driven processes, which it plausibly could be.
- Even if LLM-driven systems produce fewer bugs than human teams, courts and liability frameworks aren't equipped to handle a world where no individual "wrote" the problematic code—judges will still need someone to prosecute.
- People will adopt AI-driven systems aggressively (like Tesla FSD drivers ignoring the road) because the odds look good, then face severe legal consequences when rare failures occur, regardless of how robust their quality assurance actually was.
This is a web-based tool that generates anime art from text descriptions. You describe a character, scene, or portrait, pick a style (cinematic cel, retro 90s, chibi, pastel romance, or monochrome manga), optionally upload a reference image, and download the result — no sign-up required.
- Generates anime images in seconds from written prompts, with optional reference images to guide specific details like faces, poses, or outfits
- Offers five distinct anime styles plus control over framing (square, portrait, wide) and image resolution
- Designed for iterative refinement — users can adjust one detail at a time and compare variations to dial in their intended mood, lighting, or composition
- Allows commercial use of generated images if the source materials and references are ones you have permission to use
The author argues that LLMs don't meet the criteria that make technologies truly "inevitable" — they can't match human quality at high levels, they only replace low-value work, and they don't deliver the overwhelming performance gains that force adoption. Just because a technology is powerful doesn't mean society has no real choice about using it.
- A technology becomes inevitable only when it massively outperforms alternatives, performs socially valuable work, and leaves no way to neutralize its impact without radical behavioral change. Firearms met all three criteria in medieval Europe; the magnetic compass didn't, and remained optional.
- LLMs can only replace the least valuable parts of knowledge work — mediocre code, superficial research, low-quality design — while human expertise remains essential for anything requiring real quality. The more important the work, the less useful LLMs become.
- Real-world LLM efficiency gains are modest (20-50% improvement in low-stakes work like dashboards) and shrink dramatically as you approach actual business bottlenecks. If LLMs were truly overpowering, we'd already see tech companies producing vastly more output with them.
Researchers used BootLoops, an AI tool built with Claude, to scan scientific literature across 22 fields and identify problems that could be solved by borrowing techniques from unrelated disciplines. In 3 months, the team produced 36 papers solving problems ranging from calculating evolutionary tree probabilities to creating an 8x larger database of language sound systems.
- BootLoops discovered that many unsolved problems in different fields share identical mathematical structures with known solutions from physics, computer science, or math—but researchers didn't know to look across disciplines.
- The AI solved 15 previously unsolved integrals in high-energy physics and created ACCSTACK, a database of 6,000+ languages' stress and tone patterns that's eight times larger than the previous largest dataset.
- The approach requires heavy human judgment to work: the AI lacks "research taste" and can't determine what scientists actually care about, and it tends to favor older, highly cited work over new problems.
A new startup called Hone, founded by former employees from Cognition and Ramp, just raised $60 million to build AI agents that can handle extended business tasks over weeks or months without human intervention. The company is valued at $285 million after the funding round led by Benchmark and Index Ventures.
- Hone is developing AI agents designed to function as autonomous professional staff, capable of managing complex, long-duration tasks
- The company raised $60 million in seed funding at a $285 million valuation, backed by top-tier VCs Benchmark and Index Ventures
- The founding team comes from Cognition (an AI agent company) and Ramp (a fintech platform), bringing expertise in both AI systems and business automation
A new category of AI models that pick answers instead of generating them launched in September 2026 and immediately spawned dozens of competitors. The article cuts through the noise by laying out practical criteria for choosing which decision model to use in production rather than tracking which specific model is "best."
- Jev launched September 15 and hit massive adoption (13% of Vercel's paid customers integrated it within a day), but within two weeks Amazon, OpenAI, Cloudflare, and others shipped competing models with different tradeoffs.
- DecideBench, an independent benchmark testing 400 multiple-choice decisions across eight task families, shows Jev at 98% accuracy for $32 per million tasks, but smaller open models like Laya are 35-59% accurate and much faster—accuracy-speed tradeoffs matter depending on your use case.
- The real decision criteria are deployment location (hosted vs. self-hosted), input types (text-only vs. image support), latency requirements, cost at your actual volume, and API compatibility for future switching—not which model has the best benchmark score.
This is a news digest from The Register covering major developments across AI, infrastructure, security, and open source — including TSMC's US expansion, Microsoft's AI agent plans, Russian phishing campaigns, and various software releases.
- TSMC is partnering with GlobalFoundries on a $2B deal to boost US production of silicon interposers needed for AI datacenter chips
- Microsoft is integrating Copilot agents into Windows file systems and replacing Intel processors with Nvidia SoCs in Surface Laptop Ultra, starting at $2,599
- Russian attackers are impersonating Signal support to run phishing campaigns, and Microsoft's SharePoint on-premises servers are under active zero-day attack with unpatched vulnerabilities
- Ubuntu's new beta adds Rust components and LLM-based speech-to-text; Firefox 157 gets a significant visual redesign; KDE turns 30 with an AI-native desktop proposal
Google paused its open-source bug bounty after AI flooded it with fake vulnerability reports, forcing security teams to waste time triaging hallucinated exploits. Nulsec proposes an automated sandbox that pre-screens submissions to filter out AI slop before human engineers see them.
- Google froze its open-source bug bounty program because AI-generated reports created an unsustainable triage load—most submissions describe non-existent vulnerabilities or non-reproducible exploits, wasting hundreds of engineering hours.
- Nulsec's core mechanism: automatically execute reproduction tests in isolated sandboxes for each incoming report, blocking hallucinated claims before they reach human security engineers.
- The business model stacks three revenue streams (tiered SaaS, usage-based compute, enterprise integrations) and uses free tier access to top GitHub repos as a go-to-market wedge.
Copenhagen startup BioinvestGPT runs virtual versions of drug trials with its AI models, and it says its forecasts were right in five of six cases so far. The companies behind these tools hope that testing drugs in simulation first will help more of them succeed in real human studies.
- BioinvestGPT shared its predictions for several high-profile drug trials with Reuters in July, before those results were public.
- The firm has been correct in five of six of those forecasts so far, a small sample that still needs more track record to judge.
- The pitch is that better virtual trials could cut the failure rate of expensive human studies.
Trump issued an executive order requiring federal agencies to use "Super Intelligence" instead of "Artificial Intelligence" in official documents, and his administration is now characterizing people who use the term "AI" as enemies. The move has drawn criticism from those who see it as government overreach on language use.
- Trump's September 29 executive order mandates all federal agencies adopt "Super Intelligence" and "SI" terminology in official communications and documents.
- The White House framed the terminology shift as honoring the technology's potential and America's leadership, though the rationale for the specific rebranding remains unclear.
- Critics argue the mandate is un-American government control of speech, while others simply called it lame; the shift also aligns with Elon Musk's recent terminology changes.
@gabegreenberg on X
Gabe Greenberg replies to a post about Will Brewer, suggesting that Brewer's interest in something beyond memorizing lines may be tied to his recent sale of an AI company for around $500 million. The linked LA Times piece reports that Ben Affleck's AI company was sold for more than $500 million.
- Ben Affleck's AI company sold for more than $500 million, per the LA Times report linked in the post.
- Greenberg's reply implies Will Brewer's work goes beyond acting, pointing to his business side as the more significant story.
Anthropic released Claude Haiku 5.5, a small language model priced 75% lower than its predecessor while delivering significantly better performance on tasks like summarization, customer support, and coding. The company also cut Sonnet 5.5 cache read prices in half and added monthly API credits for paid subscribers.
- Haiku 5.5 costs $0.10 per million input tokens (vs. $1.00 for Haiku 4.5) and outperforms Haiku 4.5 across benchmarks—scoring 72.4% on OSWorld computer use tasks compared to 15.7% for the older model.
- The model includes an adjustable effort setting that lets users trade off speed for accuracy, and is positioned as ideal for high-volume, cost-sensitive work like batch summarization, database queries, and subagent tasks in larger AI systems.
- Sonnet 5.5 cache reads now cost 50% less ($0.10 instead of $0.20 per million tokens), reducing agentic task costs by ~20%, while Max and Team subscribers get monthly API credits ($100–$500) to experiment with building agents and applications.
Anthropic is expanding its Claude Startups program to give early-stage founders free access to Claude's models, API credits, and discounts on business tools. The program targets startups founded in the last five years or funded in the last two, offering up to $7,000 in Claude products plus up to $45,000 in third-party tool credits.
- Members get a free year of Claude Team (up to five Premium seats), $1,000 in API credits, and access to Anthropic's Applied AI team via office hours
- The new Claude Startup Stack bundles discounts from partner companies across sales, design, and data engineering tools
- Startups get distribution through Claude Marketplace and community access via events like Founder House and the Startup Newsletter
Anthropic released Claude as an add-on for Google Docs, Sheets, and Slides, letting you work with Claude in a sidebar without switching apps. You can also edit Google files directly from Claude using new connectors that pull in data from other sources like Salesforce.
- Claude can read, edit, and propose changes to Google Docs (with suggestion cards), build formulas and pivot tables in Sheets, and create matching slides in Slides — all without leaving the file
- You control how Claude operates: "Ask before edits" mode requires approval for each change, while "Accept all edits" mode lets Claude work autonomously
- Enterprise plans get full feature parity including Compliance API, customer-managed encryption keys, and audit export; connectors respect your existing Google sharing permissions
OpenAI released the Decisions API in public beta, which makes classification and decision-making tasks roughly 10 times faster than their Responses API. The tradeoff is higher cost per request and currently no token caching support, which some developers find economically unattractive for certain use cases.
- Decisions API processes text and image inputs to return typed answers (probabilities, choices from sets, rubric scores) about 10x faster than the Responses API, with median response times around 0.3-0.8 seconds on typical connections
- Accuracy is comparable to Luna (the model used in Responses API) but slightly faster and more variable in performance, though the practical difference is negligible
- No token caching is currently available, making the API more expensive per request than alternatives for repeated classification tasks—a limitation developers hope gets addressed soon
This is Google DeepMind's technical documentation for Nano Banana 2.1, an image generation and editing model based on Gemini 3.6 Flash that handles text and images with up to 1M token context. The card details its capabilities, known limitations, safety evaluations, and performance benchmarks.
- Nano Banana 2.1 handles text-to-image generation and image editing with a 1M token input window, 4K token image output, and 64K token text output, evaluated against public benchmarks and internal regression tests.
- The model has documented weaknesses: poor small text rendering, inconsistent character appearance between input and output, spatial confusion, and limited 3D reasoning and factuality compared to larger models.
- Safety red teaming found similar or improved performance versus Gemini 3 Flash on child safety and content policies, with no egregious concerns; frontier safety assessment indicates no tracked or critical capability level risks.
Musubi released PolicyLM-1.7B, a lightweight decision model that applies content policies written in plain English to messages in under 50 milliseconds without needing retraining when policies change. The model costs and runs as fast as traditional AI classifiers but offers the flexibility of modern language models to handle complex moderation rules.
- PolicyLM-1.7B can process content policies in plain English and make binary moderation decisions in under 50 milliseconds, matching the speed and cost of existing classifier systems while avoiding the need for retraining when policies shift.
- Decision models output predetermined outcome probabilities rather than text, making them faster and cheaper than large language models while keeping transformer architecture flexibility — a trend that started with TypeSafe AI's Jev in September, followed by OpenAI and Amazon releases.
- The model gives platform teams a scalable way to label and understand content at scale without building custom training pipelines for each policy change, addressing the challenge of moderating exponentially growing content volumes.
Sigil Wen, a Thiel Fellow who grew up in Silicon Valley's AI circles, launched Underdog, an AI assistant that runs entirely on your computer rather than in the cloud. The product makes money from transaction fees on payments the AI helps you make, not from selling your data or showing ads.
- Underdog runs a 27-billion parameter model locally on Mac and Windows (with Linux, iPhone, Android coming soon), keeping all user data on-device and encrypted, with no cloud servers involved.
- Wen claims the smaller on-device model performs comparably to Claude Opus from six months ago on benchmarks and handles everyday tasks like research and homework questions without needing state-of-the-art capabilities.
- The business model takes a small percentage of payment transactions processed through Stripe instead of charging subscriptions or mining user data, aligning the company's incentives with user privacy rather than data extraction.
Anthropic is expanding its Claude for Startups program to offer qualifying companies a free year of Claude Team (up to five seats) plus $1,000 in API credits, along with access to the Claude Marketplace and office hours with their Applied AI team. The program targets companies founded within the last five years or funded within the last two years.
- Free year of Claude Team with up to five premium seats, plus $1,000 in API credits for building with Claude
- Access to Claude Marketplace for building plugins and virtual office hours with Anthropic's Applied AI team
- Eligibility requires founding within five years or funding within two years; applications open through the Claude for Startups program page
HackerRank is rolling out Chakra, an AI agent that conducts coding interviews while candidates have access to AI tools, measuring not just whether they get the right answer but how they think and work with AI. The company claims this approach cuts cheating flags by 70-80% and consolidates a three-round hiring process into one interview.
- Chakra has already conducted 500,000+ interviews in beta with companies like Snowflake and Capgemini, and HackerRank is positioning it as the future of its business, comparing the shift to Apple moving from iPod to iPhone.
- Candidates can use AI during interviews without penalty, which HackerRank found actually reduces cheating because it removes the incentive to secretly use outside tools—suspicious activity dropped 70-80% compared to traditional assessments.
- The AI measures "AI fluency" (how well candidates prompt AI and evaluate its output) and reasoning process rather than just final answers, while HackerRank says humans retain final hiring decisions to avoid over-delegating to algorithms.
- Regulatory scrutiny of AI hiring tools is already happening (NYC requires bias audits), and while Ravisankar argues properly tuned AI is less biased than humans, automated systems can inherit biases from training data and criteria.
@liquidai on X
Liquid AI has added image input to d1, its decision model, which returns probabilities for yes/no, choice, or score questions in a single forward pass. The company says d1 matches or beats GPT-6.1 Sol on four of six real tasks while costing 19x to 200x less.
- d1 now takes images, text, or both, and was tested on six real applications including support ticket filtering and circuit board inspection.
- It matches or beats GPT-6.1 Sol on four of the six tasks, per Liquid AI's own testing.
- Text decisions come back in 200 to 300 ms, and the model is priced 19x to 200x below GPT-6.1 Sol and Claude Opus 5.5.
Bank of Japan official Uchida warned that while AI adoption has boosted demand and economic activity, a gap between inflated asset valuations and actual profits from AI investments could trigger a market pullback. The central bank is monitoring whether AI will push inflation above its 2% target and affect Japan's natural interest rate.
- AI has acted as a major demand shock supporting economic activity, but elevated stock valuations may not match the actual profits generated by AI-related investments, creating correction risk.
- Stock price gains from AI have eased financial conditions, but heavy bond issuance by AI companies has simultaneously pushed up long-term interest rates, creating conflicting pressures.
- The BOJ has identified AI-driven demand as a potential factor pushing inflation above its 2% target, which could require further monetary tightening.
Google froze its open source bug bounty program starting October 1 because AI-generated submissions overwhelmed the system with invalid reports and hallucinations. The company won't reopen it until Q1 2027.
- Google engineers and open source maintainers were flooded with automated submissions that were mostly invalid or contained AI hallucinations, making the program unmanageable.
- The pause confirms cybersecurity experts' earlier warnings that AI-generated content poses a real threat to bug bounty programs by degrading their signal-to-noise ratio.
- Google is redirecting participants to its other bug bounty programs while it figures out how to handle the AI submission problem.
Elon Musk says the company currently called SpaceXAI will be renamed SpaceXSI. The article itself is only a headline and topic stub, so the reason for the change is not in the text.
- Musk announced that the SpaceXAI name will change to SpaceXSI.
- The provided text has no details on the rationale, timing, or what the new name covers.
Hotelist is a hotel search engine that bypasses the commission-driven rankings of traditional booking sites by using AI to analyze real traveler reviews, room photos, and verified amenities instead. It's built by Pieter Levels and aims to surface genuinely good hotels rather than ones that paid for placement.
- Hotel booking sites are corrupted by pay-to-play mechanics: hotels pay higher commissions to rank higher, and review platforms are flooded with fake ratings that can be bought for hundreds of dollars
- Hotelist uses vision AI to analyze actual room photos for cleanliness and condition, and scrapes real traveler discussions across the web to extract honest pros and cons independent of hotel marketing
- Amenity claims get verified against photos—so a "gym" must actually have weights, not just one treadmill, and filters can target real coworking desks, fast wifi, and actual kitchens rather than marketing fluff
Meta released open-source code so anyone can build their own Muse AI gadgets using ESP32 boards or Raspberry Pis, and is also giving away 5,000 of its own Muse Home Link device to control smart home devices. The company is essentially letting hobbyists turn the AI agent into whatever form factor they want.
- Meta open sourced SDKs for Muse so developers can load it onto off-the-shelf hardware like E Ink displays, HDMI sticks, or touchscreens
- The company manufactured 5,000 Muse Home Link gadgets and opened a waitlist for free units shipping this month
- Muse Home Link can control lights, TVs, printers, and other devices through community-built skills
AgentCraft is a tool that runs Claude agents inside a Minecraft environment where you can watch them work on your code in real time. You give them a goal, they plan tasks, write code in isolated git branches, and ask you for decisions when needed — nothing merges without your approval.
- Agents work in their own git worktrees on isolated branches; your main checkout is never touched, and merges only happen when you explicitly approve them
- The interface visualizes agent work through a pixelated Minecraft studio where agents walk to desks, show live code changes on monitors, and come find you with questions
- A real test run on a sample repo landed six features with passing tests, handling merge conflicts, permission prompts, and a Foreman restart — all driven through the game interface
- Small goals cost $2-10 in Claude API usage; you can try it free with a simulated team first
Redis published benchmark results showing its memory system outperforms other production systems on LongMemEval, a standard test for long-term agent memory. The report also flags that published memory benchmarks often don't replicate when re-tested independently.
- Redis' Remis + Instruct strategy achieved 86.5% task-averaged accuracy on LongMemEval, the highest among production systems tested on GPT-4o, at lower cost than competitors
- Oracle's managed service scored slightly higher (within 0.5 points) but uses a stronger answer model
- Vertex Memory Bank's actual performance (42.4%) significantly underperformed its published results (~70%), suggesting benchmark results don't always hold up in real-world setups
TypeSafe AI's decision model Jev autonomously completed Pokémon Red by reading game memory, evaluating legal moves with game facts, and making every decision—from battle tactics to menu choices—over 37 hours and 16,150 decisions. The entire run cost $1.65 in API tokens, with full source code and setup instructions open on GitHub.
- Jev made every game decision (battles, catching, team swaps, navigation) based on legal options and contextual facts, while the harness only handled emulation and button presses—no cheats or memory writes.
- The project cost $1.65 total at $0.042 per million input tokens, with typical calls around 1,200 tokens and median decision time of 0.4 seconds; the full run consumed ~39.2M input tokens.
- All code is open-source (GPL-2.0) and reproducible locally with your own legally obtained Pokémon Red ROM, Node 20+, and macOS or Linux—no Nintendo assets or ROMs are distributed.
@AdrienneLaF on X
Adrienne LaFrance shares an Atlantic newsletter piece that she found funny, linking to an item about an AI slowdown framed as a "bone dissolver." The post itself is just a short reaction, so the substance lives in the linked newsletter.
- The post is a brief reaction ("This made me laugh") with no added argument, so the content is entirely in the linked Atlantic newsletter.
- The linked piece is in The Atlantic's newsletter section and is dated October 2026, with the slug pointing to an "AI slowdown" and a "bone dissolver."
- Without the full article text, I can't confirm specific claims or figures; the post alone doesn't support more detail.
OpenAI is releasing Dots, autonomous AI agents powered by GPT-6 Astra that run 24/7 to handle tasks independently or alongside you across ChatGPT, Slack, and Teams. They learn your preferences over time, connect to 4,000+ apps, and include built-in safeguards so you control what they do.
- Dots can work autonomously on their own cloud computer, access connected apps, and handle multiple projects simultaneously, or work directly alongside you on your device with permission
- The system includes granular controls: you can set custom rules for when dots act independently vs. require approval, review their work in real time, and certain sensitive actions like password changes always require your involvement
- Specialist dots for enterprises come with their own identity and credentials, integrating with Microsoft Agent 365, designed to handle defined responsibilities like procurement and customer support with team oversight
Trump issued an executive order directing federal agencies to call artificial intelligence "super intelligence" or "SI," claiming the word "artificial" makes it sound fake. The rebrand has minimal public support and confuses existing tech terminology, with experts doubting it will stick.
- Trump announced the rebrand at the UN on September 22 and signed an executive order on September 29 requiring all executive branch agencies to use "super intelligence" in official documents and communications, though he can't force private companies or foreign governments to follow suit.
- "Superintelligence" already has a specific meaning in AI research — it refers to hypothetical AI systems smarter than humans in all ways — making Trump's use of "super intelligence" misleading and confusing to the field.
- A YouGov poll found only 9 percent of the public supports the new name versus 53 percent who prefer "artificial intelligence," and multiple AI experts said the term won't catch on because it exaggerates current AI capabilities.
A researcher at Arize built a rigorous evaluation system to test whether Claude Opus 5.5 actually fixed the distinctive writing quirks that plague Claude outputs. The answer: it improved by 50%, but "fixed" overstates it — the model still relies heavily on rhetorical moves that make AI-generated text feel artificial.
- Em dashes dropped dramatically (from 12.9 to 0.035 per 1,000 words), but other Claudisms only halved across six categories: salience flags, verdict intensifiers, signposts, contrast reframes, stock metaphors, and gotcha framing.
- The real problem isn't specific phrases—it's rhetorical moves. Claude doesn't say "load-bearing" that often, but it constantly tells readers something matters instead of showing why, which no phrase list can catch.
- Evaluating writing quality requires human annotation first, then building automated checks against those annotations. The researcher manually flagged 153 instances across 53,000 words to create ground truth before training an LLM judge.
Jake Goldsborough argues that AI coding agents aren't killing software engineering—they're freeing engineers from tedious mechanical work so they can focus on the actual problem-solving that defines the craft. The hard parts that never mattered (regex wrestling, syntax memorization, build system incantations) are disappearing, but the work that requires real judgment and responsibility remains.
- The difficult parts of coding—parsing edge cases, remembering obscure syntax, navigating unfamiliar codebases—were never the point; they were tolls between understanding a problem and solving it, and removing them doesn't diminish engineering skill.
- Agents handle syntax and scaffolding faster, which lets engineers explore domains that were previously too expensive (rewriting programs in different languages, understanding unfamiliar systems, testing experimental ideas), actually tightening the learning loop rather than replacing it.
- The real shift is architectural: engineering moves from "produce every character by hand" to "direction, judgment, verification, and responsibility," and the actual risk isn't that the work disappears but that companies cut people, generated code creates failures at scale, and the junior-to-senior training path collapses.
A user searching for old jokes about NBA player Dario Saric got an AI-generated response treating the query as relationship advice instead of returning actual search results. The article argues Google has abandoned its core function of finding information in favor of forcing AI chatbot interactions into search results.
- Google's AI Overview misinterpreted a niche sports meme query as a personal relationship problem and offered emotional support instead of search results
- The actual relevant links the user wanted were buried below the AI-generated content, making Google's search function actively worse at its stated job
- The author questions whether search engines should be designed to have parasocial conversations with users rather than simply return relevant information
This is a tool that lets you animate static images by applying motion from reference videos, giving you more control over character movement and consistency than text-prompt-only AI video generation. You upload a character image, add a motion reference clip, and generate a video with free credits.
- Motion references replace guesswork: using real video clips to guide movement is more reliable than prompt-only generation for precise gestures, pacing, and body mechanics.
- Character consistency stays intact: the tool preserves the original character's appearance and identity while applying new motion, useful for avatars and repeatable content.
- Free-to-test workflow: you get free credits to iterate and compare results before committing to production, lowering the barrier to experimentation.
AI is collapsing the time needed for legal work, forcing law firms to abandon hourly billing before their revenues crater. Pryce is a SaaS tool that uses historical billing data and AI execution logs to help firms price fixed-fee engagements with confidence and protect margins.
- Corporate clients are now explicitly demanding AI discounts on invoices, making the billable hour model unsustainable for firms that adopt efficiency tools
- Law partners fear fixed-fee pricing because they lack data to accurately scope risk and quote rates—Pryce solves this by benchmarking past human hours against actual AI task completion times
- The platform builds lock-in through a proprietary dataset linking task hours to AI execution logs; as firms use it more, their firm-specific margin models become harder to replicate elsewhere
Meta released Muse Realtime Avatar, a system that generates expressive video avatars synchronized with AI speech in real time. It achieves 870ms latency while outperforming existing avatar products from Runway and HeyGen in user preference tests.
- The system reduces computational cost by 60x through distillation, condensing a teacher model's 120 diffusion steps into a student model's 2 unguided steps while preserving quality.
- Muse generates 448x768 video at 25 fps with ~870ms end-to-end latency, handling 12 concurrent sessions per GPU through optimized inference, KV cache reuse, and four-bit quantization.
- Head-to-head comparisons with leading commercial avatar systems showed raters preferred Muse Realtime Avatar across visual quality, synchronization, character consistency, and mannerisms.
@princedoesai on X
A short X post where the author says an episode of Rick and Morty made Jev AI clearer to him than a product demo would have. The post itself is a reaction rather than a detailed explanation, so the substance is mostly the claim that the show did the job of an onboarding video.
- The author says a Rick and Morty episode explained Jev AI better than any tech demo.
- The post offers no specifics about what Jev AI does or how the show framed it.
- The main signal is the author's enthusiasm, not a technical claim or finding.
Meta's new AI app Muse has been downloaded more times in its first 12 days than ChatGPT was in the same period, according to app analytics firm Apptopia. The comparison is limited to U.S. and Canada iOS downloads, but Muse also shows higher daily active users.
- Muse hit 1.8 million iOS downloads in the U.S. and Canada within 12 days versus ChatGPT's 1.3 million in the same markets and timeframe
- Muse's U.S. daily active users (642,000) significantly exceed ChatGPT's early numbers (231,000), even when narrowed to iOS only (359,000 vs. 231,000)
- Meta's integration of Muse across Facebook, Instagram, and WhatsApp—plus existing cross-promotion expertise from Threads—positions the app for rapid growth, with 95% of current Muse users already on Facebook
OpenAI released Astra for Law, a specialized AI product built on GPT-6 Astra that combines legal research capabilities with custom instructions for legal analysis and writing. Law firms can use it to build their own AI applications while maintaining client confidentiality and ethical controls.
- Astra for Law includes a legal search index and custom instructions designed to ground legal research in precise citations and authorities, rather than generating unsupported answers
- The product is designed with "legal-grade trust and controls" including information permissions, ethical walls, and client instructions, developed with input from Latham & Watkins
- It's positioned as a foundation for law firms and legal tech companies to build custom workflows around their own expertise, not a replacement for existing legal research tools like Thomson Reuters