Pulse
What's shipping in AI right now — curated, summarised, updated through the day.
- GitHub (1jehuang/jcode)·AGENTS·2w ago
jcode v0.12.4 ships as a provider-agnostic Rust coding harness
jcode hit v0.12.4 on 23 May 2026, sitting at roughly 6,500 GitHub stars across 66 releases. Open source (MIT), Rust core, supports Claude / OpenAI / Gemini / Copilot as interchangeable backends, with built-in browser automation, persistent agent memory, and multi-agent swarms. Positioned as an alternative to the Claude Code / Cursor lock-in pattern; if you've been wanting an agent harness where the model behind the curtain is genuinely swappable, jcode is the cleanest current attempt. Worth a Friday evening if you're in that bucket.
- Socket·INFRA·3w ago
Socket raises $60M at $1B for AI-era software supply chain security
Socket closed a $60M Series C at a $1B valuation on 21 May 2026, led by Thrive Capital with Andreessen Horowitz participating. The product detects malicious open-source packages before they reach production; the company claims it caught the compromised Axios package six minutes after publication. The customer list is the AI dev stack itself: Anthropic, xAI, Replit, Cursor, Figma, Vercel. As agents pull more dependencies into more codebases, automated supply-chain checks stop being defensive hygiene and start being a default-install line in any agent harness.
- OpenPR·MCP·3w ago
AVCLabs ships MCP server for AI agent video and image editing
AVCLabs launched its MCP server on 21 May 2026, exposing video enhancement and image segmentation to AI agents through natural-language calls. Notable as one of the first creative-media MCPs (most MCPs so far have been ops, dev tools, or business apps; nobody's really shipped media editing pipelines). The use case: an agent that can fetch a clip, enhance it, segment a subject, and pass the result to a video editor MCP, without leaving Claude or Cursor. Whether the quality holds up against the standard creative-tool UX is the open question.
- TechCrunch·NEWS·3w ago
Exa Labs raises $250M for agent-native search infrastructure
Exa Labs closed a $250M round at a $2.2B valuation led by Andreessen Horowitz, announced 20 May 2026. The thesis: "search for agents" is now a distinct category, separate from consumer search products like Perplexity. The same TechCrunch piece noted Parallel Web Systems closed $100M in the same window. Signal: the AI search infrastructure category has graduated from sub-bet to standalone wedge worth nine-figure rounds. Whether the economics support that many specialised search-for-agents providers is the question nobody is asking yet.
- Camunda·AGENTS·3w ago
Camunda announces ProcessOS, an agentic OS for enterprise workflows
Camunda unveiled ProcessOS at CamundaCon in Amsterdam on 20 May 2026, an AI intelligence layer sitting on its existing workflow-orchestration platform. The pitch: describe the desired outcome in natural language, four AI agents work across the lifecycle to discover how you actually run today, re-engineer for outcomes, build and deploy, then continuously optimise once live. Closed beta now, runs natively on AWS with Bedrock and AgentCore integration. Brings agentic AI to the BPMN crowd that LangChain-style frameworks haven't touched.
- GitHub (tinyhumansai/openhuman)·AGENTS·3w ago
OpenHuman v0.54.0 tops GitHub trending with a local-first memory tree
OpenHuman, an open source desktop AI agent with a SQLite-backed "Memory Tree" and 118+ OAuth integrations, hit v0.54.0 on 19 May 2026 and crossed roughly 26,000 stars on GitHub. The differentiator is its inversion of the agent cold-start problem: the app's first job is to read you, building local-first context before doing work. Rust core, Obsidian-compatible markdown memory, GPL-3.0. Plays in the same space as Reflect and mem0 but with a privacy-first stance that nudges it adjacent to Osaurus rather than to Claude Code.
- TechCrunch·INFRA·0mo ago
Osaurus launches Mac-only LLM server bridging local and cloud models
Osaurus, an open-source Apple-only LLM server from ex-Tesla/Netflix engineer Terence Pae and Sam Yoo, hit TechCrunch on 15 May 2026 after a year of public building and over 112,000 downloads. The pitch is a single Mac-native interface that can route between local models (DeepSeek v4 and similar, requires ~128GB RAM for the big ones, 64GB for smaller) and cloud providers like OpenAI and Anthropic, while keeping memory, files, and tools on your own hardware in a hardware-isolated sandbox. Differentiated from Ollama and LM Studio by going for a consumer-friendly UI rather than a developer terminal.
- Fortune·TOOLS·1mo ago
Martha Stewart-backed Hint launches AI for home maintenance
Hint, co-founded by Martha Stewart with backing from Slow Ventures, launched in mid-May 2026 with seed funding to track home maintenance, insurance, utility costs, and repairs through an AI agent. Differentiated from generic home-management apps by being agent-first rather than form-first: the AI proactively flags things like "your boiler service is overdue" rather than waiting for users to enter dates. One of the more unusual celebrity-AI launches of 2026 in that the product solves an actual recurring annoyance rather than being celebrity-merch with an AI label.
- Digiday·MCP·1mo ago
TikTok ships MCP server for agent-run ad campaigns
TikTok launched its Ads MCP server on 13 May 2026 alongside a broader developer toolkit. The server lets marketers connect AI agents directly to the TikTok ads platform so the agents can plan, launch, optimise, and report on campaigns without manual intervention. Notable as one of the first MCPs from a major ad platform; Meta's AI Connectors for Ads followed a similar pattern but with different scope. Whether agent-run paid social actually performs better than human-run remains an open question.
- Business Wire·MCP·1mo ago
Mobbin launches MCP server, 621,500 real app screens for AI agents
Mobbin officially launched its MCP server on 13 May 2026, giving AI agents access to over 621,500 indexed app screens from real shipped products. The pitch is grounding AI-generated UI in reality rather than letting agents invent design patterns from scratch. Setup is a single `claude mcp add` command pointing at api.mobbin.com/mcp, with browser-based OAuth and no API key. Available on all paid Mobbin plans, no free tier access. Most useful for designers and product teams who already pay for Mobbin and use Claude Code, Cursor, or Codex.
- PulseMCP·MCP·1mo ago
Qarera ships MCP for job search and AI-powered resumes
Qarera released its MCP server on 13 May 2026, exposing job search, AI-assisted resume generation, and application tracking through a single MCP surface. Pitched at job seekers who want to drive applications from Claude or another MCP-capable assistant rather than juggling a half-dozen tabs across LinkedIn, Indeed, and ATS portals. One of several "personal productivity" MCPs landing in May 2026 alongside Quizlar (voice-led flashcards) and Quelvio (enterprise knowledge with cited answers).
- TechCrunch·TOOLS·1mo ago
Dessn raises $6M for a design tool that works directly on production code
Dessn, founded by Gabriella Hachem and Nim Cheema, raised a $6M seed led by Connect Ventures with Betaworks and N49P participating. The pitch is design tooling that operates on your actual codebase in the cloud rather than producing Figma files that get translated by a developer later. Already in use at Color, Wispr, and Mercury. Free tier gives one compiled repo and five prompts a week; paid starts at $39 per user per month. Co-founder Cheema's framing: "in a world where code is insanely cheap, design becomes a way that's a differentiator."
- Asana Developers·MCP·1mo ago
Asana, ClickUp, Wrike, Smartsheet, Plane, Shortcut all ship official MCPs
The window from February to April 2026 saw six project-management vendors ship official MCP servers in quick succession: Asana, ClickUp, Wrike, Smartsheet, Plane, and Shortcut. The pattern across all six is the same: remote server, Streamable HTTP, OAuth via browser, no API tokens. ClickUp's surface is currently the largest at roughly 49 tools across 14 categories. Asana's is the most enterprise-friendly with workspace admin gating. For teams already standardised on one of these, the integrations are now default-install rather than experimental. For teams choosing between them, MCP surface area is now a meaningful differentiator.
- Product Hunt·AGENTS·1mo ago
Minions hits Product Hunt #4 with single-board workflow agents
Minions, an AI workflow automation tool that gives users a single task board to view all running agent tasks, launched on Product Hunt on 8 May 2026 and finished the day at #4 with 304 upvotes. The pitch is periodic check-ins, automatic retry when stuck, and escalation when alternatives are exhausted, rather than fire-and-forget agents that vanish into the void. Sits alongside the broader May 2026 wave of orchestration-first launches (AgentRail, Pixcode, Polygram) that prioritise visibility and control over raw automation.
- Salesforce Developers·MCP·1mo ago
Salesforce ships Data 360 MCP in developer preview
Salesforce released its Data 360 MCP server as an open-source GitHub repository in Developer Preview, joining the wave of enterprise vendors launching official MCPs. The pitch is connecting LLMs to Salesforce data without hitting context window limits, which has been the practical blocker for large-org Salesforce automations. Different from existing community Salesforce MCP wrappers in that this comes from Salesforce themselves and is tuned for Data 360's federated schema rather than the core CRM tables. Worth watching if you're on Data 360 already, less interesting if you're just using Sales Cloud.
- AWS News Blog·MCP·1mo ago
AWS MCP server reaches general availability
AWS announced the general availability of its official MCP server on 6 May 2026. AI coding agents (Claude Code, Cursor, Windsurf) now get secure, auditable access to AWS services through MCP without third-party wrappers. Auth, audit logging, and rate limiting are handled by AWS rather than left to developers, putting it on a similar footing to Cloudflare and Azure's existing MCP offerings. For teams building on AWS who already use agent-driven workflows, the official server removes the friction of stitching together IAM permissions for community MCPs.
- Optimizely World·MCP·1mo ago
Optimizely ships remote MCP for experimentation data
Optimizely released an official remote MCP server in early May 2026, letting anyone with Optimizely credentials point Claude, Cursor, or ChatGPT at their experimentation data through OAuth. Endpoint is exp.mcp.opal.optimizely.com/mcp. The server can query projects, list experiments, inspect performance, and modify traffic splits, but stops short of publishing live changes; you still need the UI for the final Publish click. Notable as one of the cleaner first-party MCPs in the experimentation/CRO space, which has been slow to adopt the protocol.
- Bannedbyanthropic·NEWS·1mo ago·+1 source
Anthropic users launch petition site for account ban appeals
A group of Claude users has launched bannedbyanthropic.com, a petition site calling for manual review and fair appeals processes for account bans. The site appears to be a response to what users perceive as opaque or automated moderation decisions. On Hacker News, commenters shared stories of being banned for legitimate use cases, including technical writing and research queries, with little recourse. Some noted that Anthropic's acceptable use policy is broad enough to catch edge cases, while others flagged the lack of a clear appeals pathway. The site underscores growing tension between AI safety guardrails and developer autonomy.
- Qwen·MODELS·1mo ago·+1 source
Qwen3.6-27B matches flagship coding performance at 27B parameters
Alibaba Cloud has released Qwen3.6-27B, a 27-billion-parameter dense model that claims flagship-level coding performance. The model reportedly matches or exceeds larger models on coding benchmarks whilst running efficiently on consumer hardware. Qwen positions this as bridging the gap between compact models and top-tier coding assistants, targeting developers who need strong code generation without enterprise-scale infrastructure. The release continues Qwen's strategy of competitive open-weight models at practical sizes. This matters because it suggests the performance ceiling for mid-sized models is rising faster than expected, potentially reshaping assumptions about the hardware needed for serious coding assistance.
- Reuters·NEWS·1mo ago·+1 source
Ex-CEO and ex-CFO of bankrupt AI company charged with fraud
US authorities have charged the former chief executive and chief financial officer of a now-bankrupt AI company with fraud. Reuters reported the charges on 17 April, though details of the company's identity, the nature of the alleged fraud, and the jurisdiction remain unclear from the headline alone. The case adds to a growing list of legal actions in the AI sector as regulators scrutinise funding claims and business practices. Discussion on Hacker News surfaced speculation about which company might be involved, with some pointing to recent high-profile collapses in the generative AI space. The charges matter because they signal increased enforcement attention on financial misconduct in AI startups.
- Webmatrices·INFRA·1mo ago·+1 source
Roblox cheat and AI tool caused Vercel platform outage
Vercel's hosting platform suffered a major outage triggered by an unusual combination: a Roblox cheat tool and an AI application. Details of the incident remain sparse, but the pairing suggests an exploit or resource exhaustion scenario where automated AI requests interacted with game modification software in a way that overwhelmed Vercel's infrastructure. The incident highlights the fragility of shared hosting platforms when faced with unexpected load patterns from AI tooling. Hacker News readers noted the increasingly common pattern of AI-driven traffic causing infrastructure failures, with several pointing to the challenge of rate-limiting and abuse detection for legitimate-seeming AI requests. This matters because it exposes a growing attack surface as AI tools proliferate across developer platforms.
- Reuters·NEWS·1mo ago·+1 source
NSA using Anthropic's Mythos model despite US blacklist
The US National Security Agency is reportedly using Anthropic's Mythos model, despite the company appearing on a government blacklist, according to Axios. The discrepancy raises questions about how federal procurement rules apply to AI models and whether exceptions exist for intelligence agencies. No further details on the nature of the blacklist, the NSA's use case, or Anthropic's response are available from the Reuters report. The story matters because it suggests a gap between stated policy and operational practice in government AI adoption.
- Indiatimes·AGENTS·1mo ago·+1 source
Meta tracking employee keystrokes and mouse movements for AI training
Meta is installing tracking software on US employees' computers to capture mouse movements, clicks, keystrokes, and occasional screen snapshots, according to internal memos seen by Reuters. The data will train AI models to perform work tasks autonomously, addressing areas where current models struggle, such as navigating dropdown menus and using keyboard shortcuts. Meta spokesperson Andy Stone said the data won't be used for performance reviews and that safeguards protect sensitive content. The company frames this as necessary to build agents that can complete everyday computer tasks, requiring real examples of how people actually work. It's a stark reminder that the data fuelling AI agents often comes from somewhere, and in this case, it's Meta's own workforce.
- Adweek·NEWS·1mo ago·+1 source
StackAdapt pitches ChatGPT ads based on prompt relevance
StackAdapt, an independent demand-side platform, is offering advertisers early access to place ads inside ChatGPT based on prompt relevance. According to a leaked pitch deck sent to select buyers on 27 March, the company is positioning this as a pilot programme with CPMs ranging from $15 to $60 and a $50,000 minimum spend. StackAdapt frames the offering as access to a new "discovery layer" that targets users mid-research. The deck, titled "OpenAI x StackAdapt Limited Pilot Program," describes the partnership as enabling advertising within "one of the fastest growing consumer platforms in the world." This marks a concrete step towards monetising ChatGPT's user base through contextual ad placements tied to what people are actually asking the model.
- Openclaw·TOOLS·1mo ago·+1 source
Anthropic allows OpenClaw-style Claude CLI usage again
Anthropic has told OpenClaw that reusing Claude CLI credentials for API-style integrations is now permitted, reversing an earlier stance. OpenClaw, a multi-model CLI tool, can now treat Claude CLI logins as sanctioned alongside standard API keys. The change matters for developers who already authenticate via Claude CLI on their machines and want to route requests through OpenClaw without managing separate API credentials. Anthropic API keys remain the recommended path for production workloads and clearer billing. The shift suggests Anthropic is taking a more permissive view of how developers layer tooling atop its CLI authentication, though no formal policy document has been published yet.
- Hacker News·NEWS·1mo ago
Developer fatigue with AI hype surfaces in Hacker News thread
A Hacker News discussion surfaced growing developer fatigue with AI saturation across products and job postings. The original poster compared their exhaustion to quitting Facebook, saying they'd prefer browser-level blocking of AI features. Commenters noted the technology has become 'incredibly boring', with one developer observing that engineering job specs now lead with AI branding rather than unique value propositions. The thread reflects a broader tension: whilst many developers use LLMs as productivity tools, the relentless marketing push and pressure to integrate AI everywhere is breeding resentment. This matters because developer sentiment often predicts which technologies gain lasting adoption versus which fade as overhyped trends.
- Letsdatascience·NEWS·1mo ago·+1 source
Atlassian turns on default data collection for AI training
Atlassian has enabled data collection by default across its products to train AI models, according to reports circulating in developer communities. The change appears to have been implemented without prominent user notification, prompting concern on Hacker News about opt-out mechanisms and the scope of data being harvested. Details remain sparse, with no official announcement from Atlassian clarifying which products are affected or what data is being collected. The move follows a broader industry pattern of SaaS vendors quietly adjusting terms to feed AI pipelines, raising questions about consent and transparency in enterprise tooling.
- Techcrunch·NEWS·1mo ago·+1 source
Anthropic takes $5B from Amazon, commits $100B cloud spend
Amazon is investing another $5 billion in Anthropic, bringing its total stake to $13 billion. In return, Anthropic has committed to spend over $100 billion on AWS over the next decade, securing up to 5 GW of compute capacity for training and running Claude. The deal centres on Amazon's custom silicon: Graviton CPUs and Trainium AI accelerators, covering Trainium2 through the not-yet-released Trainium4. The structure mirrors Amazon's February deal with OpenAI, where a $50 billion investment was partly delivered as cloud credits. The arrangement arrives as VCs reportedly circle Anthropic with offers valuing the company at $800 billion or more, suggesting a fresh funding round may be imminent.
- Techcrunch·NEWS·1mo ago·+1 source
Deezer: 44% of daily uploads now AI-generated music
Deezer reported that AI-generated tracks now make up 44% of all new music uploaded to its platform, roughly 75,000 tracks per day. That's up from 10,000 daily in January 2025, when the company launched its AI detection tool. Consumption remains low at 1-3% of total streams, and 85% of those streams are flagged as fraudulent and demonetised. Deezer automatically excludes AI tracks from algorithmic recommendations and editorial playlists, and will no longer store hi-res versions. The announcement follows an AI track topping iTunes charts in five countries last week. A November survey found 97% of listeners couldn't distinguish AI music from human-made tracks, and 80% wanted clear labelling. This matters because streaming fraud is scaling faster than platforms anticipated.
- Reuters·NEWS·1mo ago·+1 source
Meta to capture employee keystrokes and mouse movements for AI training
Meta plans to begin recording employee mouse movements and keystrokes to train AI systems, according to Reuters. The company has not disclosed which AI models will be trained on this data, nor whether employees can opt out. The move follows broader industry trends of using internal productivity data to improve AI assistants and code completion tools, though the scope of Meta's data collection appears unusually granular. The announcement has sparked discussion on Hacker News, where commenters questioned the privacy implications and whether such monitoring would affect employee behaviour. This matters because it signals how far companies may go to source training data for workplace AI tools.
- Axios·NEWS·1mo ago·+1 source
NSA using Anthropic's Mythos despite Pentagon blacklist
The National Security Agency is using Anthropic's Mythos Preview model, even as the Department of Defense maintains the company is a supply chain risk and pursues legal action to cut it off. The Pentagon moved in February to blacklist Anthropic and force vendors to follow suit, yet the NSA and potentially other DoD units now have access to Mythos, one of roughly 40 organisations granted use of the model due to its offensive cyber capabilities. Anthropic CEO Dario Amodei met White House officials on Friday to discuss government use of Mythos. The contradiction highlights a rift between the Pentagon's contractual dispute with Anthropic over acceptable use policies and the intelligence community's need for cutting-edge security tools.
- Claude·NEWS·1mo ago·+1 source
Anthropic removes Claude Code from Pro plan
Anthropic has quietly restructured its Claude pricing tiers, removing Claude Code from the $20/month Pro plan. The feature, which allows users to create and execute code directly in the interface, is now exclusive to the Max plan (starting at $100/month) and Team Premium seats ($100/seat/month). Pro subscribers retain access to Claude Cowork, projects, and Research, but lose the ability to run code. The move appears to push power users toward significantly more expensive tiers. Hacker News discussion suggests the change caught existing subscribers off guard, with some questioning whether the Pro tier still justifies its price point without code execution capabilities.
- Stephvee·NEWS·1mo ago·+1 source
AI Resistance Is Growing
Stephvee reports on mounting pushback against AI adoption, though without access to the full article, the specific forms of resistance and their scale remain unclear. The headline suggests organised or widespread opposition rather than isolated incidents. On Hacker News, commenters debated whether resistance stems from legitimate concerns about job displacement, quality issues, or cultural backlash against hype. The discussion touched on creative industries, academic integrity, and whether current AI tools deliver enough value to justify their costs. The story matters because developer tooling and agent adoption depend on user trust, and sustained resistance could reshape deployment priorities.
- Bsky·NEWS·1mo ago·+2 sources
Anthropic removes Claude Code from Pro subscription tier
Anthropic has quietly removed Claude Code from its $20-per-month Pro subscription, according to its pricing page. The feature, which allowed users to run code directly within Claude conversations, now appears to be absent from the Pro tier offering. Ed Zitron flagged the change on Bluesky, seeking confirmation from current subscribers. Hacker News discussion suggests the move may push users toward higher-priced tiers or separate billing for code execution features. The change comes without public announcement, leaving existing Pro subscribers uncertain about whether they retain access or face a downgrade. This matters because it signals a potential shift in how Anthropic packages developer-focused features, possibly separating code execution from general chat capabilities.
- anthropics·TOOLS·1mo ago
Claude Code v2.1.112 fixes Opus 4 availability issue in auto mode
Anthropic released Claude Code v2.1.112, a patch fixing an error where users saw "claude-opus-4-7 is temporarily unavailable" when running the tool in auto mode. The release addresses a blocking issue that prevented the model from being accessed through the interface. Claude Code is Anthropic's developer tool for AI-assisted coding, and this update restores full functionality for users relying on the Opus 4 model. The fix matters because it unblocks developers who depend on Opus 4's capabilities for automated coding workflows.
- Simon Willison·MODELS·1mo ago
Claude Opus 4.7 system prompt: less verbose, more autonomous
Anthropic updated the Claude Opus system prompt between version 4.6 (February 2026) and 4.7 (April 2026). Simon Willison extracted the changes using Claude Code to build a Git history of the prompts. Key shifts: Claude now defaults to acting rather than asking clarifying questions, uses a new tool_search mechanism to check for deferred capabilities before claiming ignorance, and has explicit instructions to keep responses concise. The child safety section was expanded and wrapped in critical tags. Claude in PowerPoint joins the list of available tools. The update also removes instructions about avoiding asterisk-based emotes, suggesting the model no longer needs that guardrail. This matters because system prompt changes reveal how labs are steering model behaviour in production.
- Hacker News·NEWS·1mo ago
Developer hit with €54k bill after Firebase key used for Gemini API abuse
A developer reported a €54,000 billing spike over 13 hours after their unrestricted Firebase browser key was used to access Gemini APIs. The incident highlights a shift in Google's API security model. For over a decade, Google told developers that API keys for services like Maps and Firebase weren't secrets and could be safely embedded in client-side code. With Gemini's introduction, that guidance no longer holds. The keys can now access expensive generative AI endpoints, turning what was once a low-risk practice into a potential financial disaster. The case underscores the need for developers to audit existing projects and apply strict API restrictions, especially when legacy keys suddenly gain access to costly new services.
- openai·TOOLS·1mo ago
OpenAI releases Codex 0.122.0-alpha.8 pre-release
OpenAI has published version 0.122.0-alpha.8 of Codex, marked as a pre-release. The GitHub release page offers minimal detail beyond the version number and commit hash (3c1a6d9), with 22 commits to main since the tag. No changelog or feature notes are visible in the release description. The release includes 93 assets, though their contents aren't specified. This appears to be a routine alpha build in an ongoing development cycle. Without release notes, it's unclear what changed or whether this signals new capabilities worth testing.
- openai·TOOLS·1mo ago
OpenAI Codex 0.122.0-alpha.6 released
OpenAI has pushed a pre-release version of Codex, tagged rust-v0.122.0-alpha.6, to GitHub. The release includes 93 assets and sits 44 commits behind the main branch. No changelog or feature notes are visible in the release page excerpt, which suggests this is an incremental alpha build rather than a major feature drop. The Rust tag prefix indicates this may be a Rust-specific build or binding. Without detailed release notes, it's unclear what's changed since the previous alpha. Worth watching if you're tracking Codex development, but not much signal here yet.
- Hacker News·INFRA·1mo ago·+1 source
Zero-copy GPU inference from WebAssembly on Apple Silicon
A developer has demonstrated that WebAssembly modules can share memory directly with the GPU on Apple Silicon, eliminating the usual serialization overhead. The technique chains three components: mmap for page-aligned memory, Metal's bytesNoCopy API to wrap that pointer without copying, and Wasmtime's custom memory allocator to use the same region as Wasm linear memory. The result is that a Wasm guest and the GPU read and write the same physical bytes, with no intermediate buffers. The developer measured zero RSS delta compared to 16.78 MB for the copy path. This matters because it turns Wasm into a viable control plane for stateful AI inference on Apple hardware, with near-zero overhead between the sandbox and the accelerator.
- openai·TOOLS·1mo ago
OpenAI releases Codex 0.122.0-alpha.12 pre-release
OpenAI has published a pre-release version of Codex tagged as rust-v0.122.0-alpha.12. The release includes 93 assets and follows three commits to the main branch since the previous version. No changelog or feature details are visible in the GitHub release notes, which appear to have encountered loading errors. The Rust-specific versioning suggests this build targets Rust language bindings or tooling. Given the alpha designation and lack of documentation, this appears to be an internal or experimental build rather than a production-ready update. Developers working with OpenAI's Codex SDK should wait for stable releases or consult the repository directly for context.
- openai·MODELS·1mo ago
OpenAI releases Codex 0.122.0-alpha.7
OpenAI has published a pre-release version of Codex tagged rust-v0.122.0-alpha.7. The GitHub release page offers minimal detail beyond the version number and 93 bundled assets. No changelog or feature notes are visible in the excerpt, making it unclear what's changed since the previous alpha. The release appears to be part of an ongoing Rust-based development track for Codex, OpenAI's code generation model. Given the alpha tag and sparse documentation, this looks like an internal or experimental build rather than a production-ready update. Developers tracking Codex's evolution may want to check the full release notes or commit history for specifics.
- Hacker News·TOOLS·1mo ago·+1 source
Claude Design and the shift from Figma back to code
A designer argues that Figma's complexity has made it irrelevant in the age of AI agents. Because Figma's proprietary format wasn't in LLM training data, models learned code instead. As agents make coding easier for designers, the source of truth is migrating back to code. The post walks through Figma's own design system files to show how baroque the tooling has become: 946 colour variables, nested modes, aliased references, and components with names like "Default, Focused, Close Button=True". The author suggests design tooling will fork into two paths as the industry reckons with this shift.
- Hacker News·NEWS·1mo ago·+1 source
Cornell instructor uses typewriters to sidestep AI in language classes
Grit Matthias Phelps, a German language instructor at Cornell, has students complete one assignment per semester on manual typewriters. The practice began in spring 2023 after she grew frustrated with students using generative AI and translation tools to produce grammatically perfect work. Students type without screens, spellcheckers, or delete keys, forcing them to slow down and engage directly with the material. The typewriter exercise is part of a broader shift in US education toward pen-and-paper exams and oral tests as instructors grapple with AI-assisted coursework. It matters because it shows one pragmatic, low-tech response to a problem that developer tools and detection software have largely failed to solve.
- Hacker News·TOOLS·1mo ago
Is It Agent Ready? A scanner for AI agent compatibility
A new tool lets you scan any website to check how well it supports AI agents. The scanner tests for emerging standards including robots.txt rules for AI bots, Markdown content negotiation, MCP Server Cards, Agent Skills, OAuth discovery, and agentic commerce protocols like x402. It covers five categories: discoverability, content accessibility, bot access control, protocol discovery, and commerce. The tool suggests quick wins like publishing a valid robots.txt with AI bot directives and exposing discovery headers. It's backed by Cloudflare's agent documentation. The scanner matters because it turns abstract agent readiness into concrete, actionable checks that developers can actually fix.
- Anthropic·INFRA·1mo ago
Anthropic orders gigawatts of TPU capacity from Google and Broadcom
Anthropic has signed a deal with Google and Broadcom for multiple gigawatts of next-generation TPU capacity, expected online from 2027. The compute will support Claude's frontier models as the company's run-rate revenue hits $30 billion, up from $9 billion at end-2025. Over 1,000 business customers now spend more than $1 million annually, double the figure from February. Most capacity will be sited in the US. Anthropic trains Claude across AWS Trainium, Google TPUs, and NVIDIA GPUs, maintaining AWS as primary partner while expanding its multi-cloud strategy. The scale of this order reflects genuine demand, not speculation.
- Anthropic·NEWS·1mo ago
Anthropic signs AI safety MOU with Australian government
Anthropic has signed a Memorandum of Understanding with the Australian government to collaborate on AI safety research and support Australia's National AI Plan. The agreement includes working with Australia's AI Safety Institute on model evaluations and risk assessments, mirroring arrangements with safety institutes in the US, UK, and Japan. Anthropic will share Economic Index data to track AI adoption across key Australian sectors including natural resources, agriculture, and healthcare. The company is also investing AUD$3 million in Claude API credits to four Australian research institutions working on disease diagnosis, genomics, and computer science education. Anthropic is exploring data centre infrastructure investments in Australia. This matters because it extends formal government-industry AI safety collaboration to the Asia-Pacific region.
- Anthropic·NEWS·1mo ago
Anthropic commits $100m to Claude Partner Network
Anthropic is launching the Claude Partner Network with an initial $100 million investment for 2026. The programme supports consultancies, professional services firms, and specialist AI companies helping enterprises deploy Claude. Partners get access to new technical certifications (starting with Claude Certified Architect, Foundations), dedicated Applied AI engineers for live deals, training materials, and co-marketing support. Anthropic is scaling its partner-facing team fivefold and releasing a Code Modernization starter kit for legacy codebase migration. The move positions Anthropic as the only frontier AI provider available across AWS, Google Cloud, and Microsoft. This matters because enterprise AI adoption depends heavily on implementation partners who handle deployment, compliance, and change management inside large organisations.
- Anthropic·NEWS·1mo ago
Anthropic launches The Anthropic Institute to study AI's societal impact
Anthropic has launched The Anthropic Institute, a research unit focused on how powerful AI will reshape society. Led by co-founder Jack Clark, the Institute combines three existing teams: Frontier Red Team (stress-testing AI limits), Societal Impacts (real-world usage), and Economic Research (jobs and economy). It will also study AI progress forecasting and legal system interactions. The Institute has hired Matt Botvinick from Google DeepMind to lead work on AI and the rule of law, and economist Anton Korinek to study how transformative AI could reshape economic activity. The goal is to share what Anthropic learns from building frontier systems and engage with communities facing AI-driven change. This matters because Anthropic believes extremely powerful AI is arriving faster than most expect.