Perplexity is adding Brain to Computer, its agent for getting tasks done. the point is to make Computer less stateless: Brain builds a context graph from past sessions, sources, files, decisions, and corrections, then uses that history to start future tasks with more relevant context.

the important bit is the shift from "memory about you" to memory about the work. Perplexity says early tests on tasks needing past context showed higher correctness and recall, plus lower cost; for now it's a research preview for Max and Enterprise Max subscribers.

i've not tried perplexity in a long time other than for searching similar products; but might give computer a try again and see how this fairs with much larger context

🔗 https://www.perplexity.ai/hub/blog/self-improving-memory-for-agents
🔗 https://x.com/perplexity_ai/status/2067642139014742348

Google is making AI avatars in Google Vids available at no cost for anyone with a Google account in the US, with other regions rolling out later. the new Vids updates turn Slides into narrated videos, add avatars and voiceovers in 24 languages, and support branded avatars for product demos.

the interesting bit is the packaging: Google is moving avatar video from a production tool into the regular Workspace flow, so sales decks, onboarding docs, and internal updates can become short presenter-led videos without setting up a shoot.

🔗 https://x.com/GoogleWorkspace/status/2067249947993366654

Amp added custom agents through plugins.

plugins can now create agents, run them once, keep talking to their threads, and even send messages between agents. useful direction for devtools: agents stop being one-off chat boxes and become programmable workers inside the tool.

🔗 https://ampcode.com/news/custom-agents

Codex Record & Replay has started rolling out on macOS in supported markets.

it lets you record a workflow once, then turn that demo into an inspectable skill Codex can reuse later. Nick’s note says EEA, UK, and Switzerland are not in the first rollout, and that computer use needs to be enabled.

🔗 https://x.com/nickbaumann_/status/2068030706077634604

Cloudflare added temporary accounts for AI agents.

with wrangler deploy --temporary, an agent can deploy to Cloudflare without a user signing up or logging in first. that’s a neat unlock for agentic deploy flows: try the thing first, attach the real account later.

🔗 https://blog.cloudflare.com/temporary-accounts/

Dot Matrix is a tiny site for generating loading animations.

it has 55+ free, open-source loaders, with controls for size, shape, animation, hover effects, bloom, and more. useful little design/dev utility when you need a polished loader without building one from scratch.

🔗 https://dotmatrix.zzzzshawn.cloud/

Matte is a native Mac app for making polished app demo videos. it records your iOS simulator, live iPhone or iPad, app window, screen, camera, and mic, then lets you edit the demo with device frames, zooms, captions, taps, swipes, overlays, and exports.

Jose’s demo shows where it is going next: a work-in-progress 3D system with realtime 60fps phone renders, ACES tonemapping, motion blur, exposure controls, and editable backgrounds inside the same editor.

🔗 https://matte.app/
🔗 https://x.com/josesaezmerino/status/2068034612912066642

Ramp has a good framing for AI spend: don't manage it by tokens, manage it by units of work.

The practical bit is to make cheap defaults the norm for routine jobs, then explicitly escalate for high-value or ambiguous work. Spend less on the autocomplete stuff so you can spend more where a smarter model actually changes the answer.

🔗 https://engineering.ramp.com/post/ai-spend-value

in the weights is a tiny leaderboard for AI model memory: type a name, it asks a bunch of models what they know, clusters the answers, and gives you a "strength" score for how much you seem to exist inside them.

i tried my handle and got software engineer, 264 strength / top 20%. not sure whether to be flattered or start doing more weightlifting 💀

🔗 https://intheweights.com/

discover.me launched an agent-built profile directory for people whose work does not fit neatly into a resume.

the idea: point Codex, Claude Code, Cursor, Cline, or OpenCode at it, and your agent reads your repos and projects, verifies proof like GitHub/domains, then turns that into a profile humans and other agents can search.

good or fluff?

🔗 https://x.com/jamiepine/status/2068472579841720584

Brighter is selling a 60,000-lumen lamp that tries to make indoor rooms feel like daylight, not just a tiny SAD light. it is dimmable from 2,500 to 60,000 lumens, shifts from 2200K to 6500K, and works with Matter smart home setups.

all the tech companies above buy their lights lol and pretty cool... 1 light to light up an entire room

🔗 https://getbrighter.com/

Ports is a small free macOS menu bar app for the kind of dev-machine mess that builds up when you have Next, Vite, Python, Rails, Go, Bun or Deno servers running everywhere.

it lists every listening localhost port with uptime, CPU, memory and an energy badge, plus quick actions to jump back to the owning terminal or kill the process. simple, but useful if your day involves a lot of local servers.

🔗 https://www.ports-app.com/

Devouring Details is an interactive reference manual from Rauno for designers who obsess over how software feels: timing, motion, intent, gestures, layout, and the small interaction choices that make an interface click.

it is not positioned as a linear course. it has 23 chapters and 23 downloadable React components on a custom platform, so you can study the thinking and pull apart the actual examples instead of just watching a tutorial.

🔗 https://devouringdetails.com/

Sakana AI launched Sakana Fugu, a model API that turns a multi-agent system into one OpenAI-compatible endpoint. instead of making teams wire up routers, specialist agents, and synthesis logic themselves, Fugu handles model selection and delegation behind the scenes.

there are two models: Fugu for lower-latency everyday work, and Fugu Ultra for harder multi-step tasks. Sakana says Ultra reaches frontier-style benchmark results, but the more useful signal is the product shape: orchestration is being sold as the model, not as app-side plumbing.

Folks are saying this is a mythos-level model. thoughts? 🤔

🔗 https://x.com/SakanaAILabs/status/2068861630327443966

Stripe introduced Stripe Directory, a preview that lets people and agents search for businesses on the Stripe network from the CLI.

the interesting bit is that it is more than a directory. results can include Stripe Apps, Stripe Projects providers, machine-payment endpoints, and Link support, so an agent can discover a service and keep moving toward integration or payment without hopping through a bunch of dashboards.

🔗 https://docs.stripe.com/directory

mmm.page says it has crossed 100k users, five years into building a looser kind of website builder.

the fun part is that it still feels more like an internet canvas than a template machine: text, images, shapes, GIFs, video, code, drawing, and no-login browsing. the founder says 2.0 is coming soon.

🔗 https://x.com/xhfloz/status/2069105854691868993

Anthropic launched Claude Tag, a Slack-based way to bring Claude into a shared team channel instead of a private chat. you can tag @Claude on a thread, give it access to selected channels, tools, data, or codebases, and let it work through tasks like pull requests, analysis, incident response, and follow-ups where the team can see the handoff.

it is beta-only for Claude Team and Enterprise for now, which is the important caveat. the bigger shift is that Claude is moving closer to a governed workplace agent: scoped access, channel memory, spend limits, and its own identity instead of one person copy-pasting context around.

🔗 https://www.anthropic.com/news/introducing-claude-tag

Aside launched a macOS AI browser that is trying to make the browser itself the agent, not just another chat sidebar. it works across websites, accounts, and history, can use your Claude or ChatGPT subscription, and leans hard on local memory plus human approval for sensitive actions.

the benchmark claims are from Aside's own run artifacts, so i'd treat them as a launch signal rather than a neutral verdict. still, it points to where this category is going: browsers built around full web-task automation, not just smarter tabs.

🔗 https://aside.com
🔗 https://github.com/at-inc/aside-benchmarks
🔗 https://x.com/hyojun_at/status/2069497198879048131

Baidu has open-sourced Unlimited OCR, a document OCR model meant to read long PDFs in one pass instead of chopping the job into page-by-page calls. the interesting part is Reference Sliding Window Attention: it keeps the visual reference plus recent output in memory, while the KV cache stays bounded as the transcript gets longer.

it is still Baidu's own technical report, but the code, model card, and paper are public. the reported result is 93.92% on OmniDocBench v1.6, with a 3B MoE model that only activates about 500M params at inference.

🔗 https://github.com/baidu/Unlimited-OCR
🔗 https://huggingface.co/baidu/Unlimited-OCR
🔗 https://arxiv.org/abs/2606.23050

Shelf is a tiny browser extension for people who already use Telegram as their read-later pile. click once in Chrome or Safari and it sends the page title, URL, domain, and selected text to a private Telegram bot chat, then lets you search, archive, and export saved links later.

the useful bit is the constraint: it only reads the page you choose when you click it, not your full browsing history. there’s a 3-day trial, then $8 every 3 months if it sticks.

(or maybe just copy and paste 💀 so you dont pay $ for apps.)

🔗 https://useshelf.dev/