Aspect is trying to be the shared file system for video teams: mount a project like a drive, open huge media in Finder without pulling the whole file down, and keep Premiere, Resolve, and Final Cut work synced for everyone else.

the interesting part is the bundle: cloud storage, local streaming, review and approval, sharing, archiving, plus an AI agent that can search and organize the media library. the "kill Dropbox" line is loud, but the real target is the messy stack around Dropbox, LucidLink, Frame.io, transfer tools, and asset managers.

🔗 https://x.com/gurishsharma/status/2064399948347568638

River AI is the new company from Igor Babuschkin, a former xAI co-founder who also worked at DeepMind and OpenAI.

the pitch is personal AI owned and shaped by the user. the first shared piece is an API preview for LoRA fine-tuning and reinforcement learning on open models, with a longer-term plan around personalization, continual learning, and local AI hardware. the founder context makes it worth watching, but it is still early until River shows the product beyond the training layer.

🔗 https://river.ai/introducing-river-ai.html

Mastercard’s new Agent Pay for Machines sounds like an AI payments launch, but the sharper read is that it’s trying to own the boring control layer before agent commerce becomes real. identity, spend limits, settlement, liability, credentials - all the stuff that has to exist before you let software start moving money on its own.

the partner list is the tell: Stripe, Adyen, Coinbase, Cloudflare, Polygon, Ripple, Solana Foundation, OKX, Aave, Lovable and others are already around the table. still early, and not exactly something normal people can use tomorrow, but Mastercard is clearly trying to become the toll booth for machine-to-machine payments before the machines show up in force.

🔗 https://www.mastercard.com/us/en/news-and-trends/press/2026/june/mastercard-launches-agent-pay-for-machines.html

Niteshift launched a full-stack cloud for coding agents and announced a $7M seed led by Greylock. teams define their app environment, services, auth, seeded data, tools, and policies once, then run Claude Code, Codex, OpenCode, Pi, or whatever agent they want against that setup.

the sharper point is verification: agents are getting good enough to run for hours, but teams still need real app stacks, tests, browser checks, logs, CI, and evidence attached to PRs before they can trust the work.

🔗 https://niteshift.dev/blog/introducing-niteshift

Recursive published early results from an automated ai research system that loops through proposing ideas, implementing them, running experiments, and validating the results. on small-model training and gpu-kernel benchmarks, it says the system improved NanoChat BPB, cut a NanoGPT Speedrun run from about 80.6s to 77.3s on the same H100 setup, and open-sourced artifacts from the work.

the interesting caveat is that this works best where the goal is clear and experiments are fast to score. Recursive also says reward hacking kept showing up, so the real story is not "ai scientist solved research", but that agents are getting better at grinding measurable parts of the ai stack when the evaluator is strong enough.

excited for the future!

🔗 https://www.recursive.com/articles/first-steps-toward-automated-ai-research

Castform is launching in open beta, with a platform for post-training open-weights models on your own data. the pitch is that you bring agent traces, docs, or reward definitions, and Castform handles data prep, gpu setup, algorithm defaults, and a training ide to watch the model learn.

the useful angle is less "everyone should train a model now" and more that fine-tuning is being productized for ai builders who are hitting api bills and latency walls. the cost/performance claims are still Castform's own framing, but the direction is clear: smaller task-specific models are becoming a real alternative to endlessly prompting bigger closed models.

🔗 https://castform.com/blog/beta-launch/ / https://x.com/googrish/status/2065107744215171486

Tigris is trying to make object storage feel more like git for agent workflows.

StorageSDK is a TypeScript SDK that lets you snapshot a bucket, fork it for each agent run, write into isolated branches, then roll back or throw them away without touching the parent dataset. the idea is simple: agents need scratch space with history, not just PUT, GET, LIST, DELETE.

Pretty cool going have to try this out on my pi

🔗 https://www.tigrisdata.com/blog/storagesdk/ | https://x.com/TigrisData/status/2065111360116023578

Pool, the screenshot app, is now live on the App Store. the pitch is simple: turn screenshots into searchable, shareable pools, find original links, and use your camera roll as a kind of personal context layer. it's free on iPhone, but the listing still says access is rolling out gradually.

the privacy bit matters here: Pool says search runs locally with bundled CLIP models, while some smarter "intent" actions are analyzed by Gemini in the cloud. App Store privacy labels also show photos/videos and other account/app data may be handled, so this is less "just screenshots" than it sounds.

🔗 https://x.com/_poolday_/status/2065097298946777226

OpenAI is letting Codex users bank rate-limit resets instead of using them only when the timer rolls over. Go, Plus, Pro, and Business users start with one free reset, and Plus/Pro users can earn another one through a two-week referral promo when a friend sends their first Codex message.

small quality-of-life change, but useful for longer coding sessions: the reset becomes something you can save for when you actually need a full run, not just whenever the quota clock refreshes.

LETS FREAKING GOOOOO

🔗 https://developers.openai.com/codex/pricing + https://x.com/OpenAI/status/2065225362544726371

recent.design is back: a simple daily index of good design work spotted on X, with the original posts linked back to the designers.

the interesting bit is the workflow behind it. Daryl Ginn says he skipped an admin panel entirely: drop a tweet into Telegram, let Claude analyze and tag it through Cloudflare Workers/webhooks, then publish after a confirmation. small, useful example of agents as publishing ops, not a chatbot.

🔗 https://x.com/darylginn/status/2065344827605311554

Moonshot released Kimi-K2.7-Code, a new open-source coding model built on K2.6. the useful bit is not just higher benchmark scores: they are pitching it as better at long coding tasks while using about 30% fewer reasoning tokens, which matters if coding agents are going to run for hours instead of one prompt at a time.

weights and code are on Hugging Face, and it is available through Kimi API and Kimi Code. the 6x high-speed mode is still only listed as coming soon.

🔗 https://x.com/Kimi_Moonshot/status/2065377579130142937

Anthropic says the US government ordered it on June 12 to suspend access to Claude Fable 5 and Claude Mythos 5 for any foreign national, including foreign-national Anthropic employees. because Anthropic can’t enforce that cleanly across customers on short notice, it says it has disabled both models for everyone. other Claude models are unaffected.

the weird part is the stated trigger: Anthropic says the government gave no detailed national-security explanation, but appears to be reacting to a narrow Fable 5 jailbreak that found minor, already-known vulnerabilities. Anthropic says other public models can find the same kind of issues without a bypass, so this looks less like a normal safety recall and more like the first serious export-control fight over frontier model access.

🔗 https://www.anthropic.com/news/fable-mythos-access

Ramp released Ramp SWE-Bench, a private coding benchmark built from real Ramp PRs that went through engineer review. it is basically SWE-Bench pointed at one company's production code instead of public GitHub issues, which makes it more useful for choosing coding agents on the work they will actually do.

the takeaway is not a clean winner. Claude Fable 5 solved the most tasks in v1, while GPT-5.5 and Opus 4.8 showed different cost and speed tradeoffs. the bigger signal is that teams are starting to grade coding agents on their own code, not public leaderboards.

🔗 https://labs.ramp.com/swebench

Z.ai has pushed GLM-5.2 into its GLM Coding Plan for Lite, Pro, Max, and Team users. it is pitching this as its new flagship coding model, with 1m context support and High/Max thinking-effort modes for deeper coding runs.

the caveat is availability: this is live inside the Coding Plan now, while API/chatbot access and MIT-licensed open weights are still promised for the week of june 15. useful if you already use Z.ai's coding stack, but the open-weight part is not here yet.

🔗 https://x.com/Zai_org/status/2065704919299235870

Newt is a pixel-art editor with a useful twist: you paint like a normal raster tool, but the project underneath is code. each pixel is tied to a color token, so palette changes can update the whole sprite instead of flattening into a bitmap.

the practical bit is export. it can turn the same sprite into SVG, CSS, React, JSON, PNG, GIF, icons, sprite sheets, or CSS keyframes, which makes it feel more like a tiny asset pipeline for apps and games than just a drawing toy.

mr qi shared this with me today and i thought it was pretty darn cool. you can even do animations directly, very very cool and minimal!

🔗 https://newt.sh

Databricks open sourced Omnigent, an alpha meta-harness for running agents like Claude Code, Codex, Pi, and custom agents through one common layer.

the useful part is that it treats agent sessions more like shared infrastructure: same session across terminal, web, desktop, and phone; live sharing with teammates; and policies for spend, permissions, sandboxing, and approvals that sit above the individual harness. less "one more agent", more a control plane for agent sprawl.

🔗 https://www.databricks.com/blog/introducing-omnigent-meta-harness-combine-control-and-share-your-agents

OpenRouter launched Fusion, a server-side compound model you can call as the openrouter/fusion model slug or as a tool. it fans a prompt out to several models, uses a judge to compare consensus, contradictions, and blind spots, then returns one answer.

the benchmark claim is interesting but narrow: in OpenRouter's own DRACO deep-research run, Fusion panels beat the solo models they tested, and a cheaper panel came close to Claude Fable 5. caveat: this is one benchmark, Fable only completed 93/100 tasks, and OpenRouter says long-horizon tasks still need testing.

🔗 https://openrouter.ai/blog/announcements/fusion-beats-frontier/

Easel is a new open source macOS app that gives Codex a visual product-design workspace. it combines a local project library, reusable design systems, Codex chat, a live web preview, and a click inspector so you can point at a UI element and ask for changes.

the interesting bit is that it stays close to the actual project files: prototypes, slide decks, assets, and design systems live locally, while Easel wraps the frontend iteration loop in a native SwiftUI app. the first release is already on GitHub with a signed/notarized DMG.

🔗 https://github.com/jamesrochabrun/Easel

an arXiv preprint from Stanford, NYU, MIT and Princeton researchers found a useful warning sign for everyday AI use: people often think AI saves more time than it does on simple tasks like arithmetic, spelling, recall and short rewrites.

across 3 preregistered studies with 2,691 people, participants expected AI to save 55.7 seconds on average, but the measured saving was 7.5 seconds. the bigger finding is behavioral: after using AI on just two tasks, people became more likely to use it again, even when doing the task themselves was faster. the risk is not that AI makes people incapable overnight, but that it can quietly blur the sense of when your own brain is already the quicker tool.

in short: how often are we using AI because it is faster, versus because it makes the work feel lighter?

🔗 https://arxiv.org/abs/2605.22687

Google Research is backing a UC San Diego experiment to turn retired Pixel phones into low-carbon cloud compute. Jeff Dean's framing makes the bet clearer: people replace phones every ~4 years, leaving hundreds of millions of still-usable devices whose manufacturing footprint already exists.

the plan is to remove the batteries and other components, power the boards directly, replace Android's userspace with a general Linux distro, and run 25-50-device clusters under Kubernetes. the first deployment is a planned 2,000-phone UCSD cluster for classes and research, expected in fall 2026.

mobile-phone level compute clusters lol, super cool

🔗 https://research.google/blog/a-low-carbon-computing-platform-from-your-retired-phones/ / https://x.com/JeffDean/status/2065649717573505188

Notra launched image generation for teams that want marketing visuals from the product work they already shipped.

instead of starting from a blank prompt, it uses shipped work, brand style, and real components, then exports into Paper or Figma. useful direction: the same workflow that writes launch posts is now making the visuals around them too.

think visual work for your PRs to show what was done!

🔗 https://www.producthunt.com/products/notra-3 / https://x.com/usenotra/status/2066415660268458204