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/

Hano is a new browser tool for making 3D device mockups and launch visuals. upload a screenshot or video, pick a device, tune the scene, then export an image or animation without setting up a 3D app.

the interesting part is the target user: designers and indie builders who need polished product shots fast. the site is live, and its own app text points to 2K/4K image export and video export as part of the product flow.

this is so freaking cool I’m gna try it for my next app 🫪🥵🙌🏻

🔗 https://www.hano.so/
🔗 https://x.com/shaigexp/status/2069501529258680438

next up:

Figma’s agent can now build generative plugins in open beta, turning a prompt into a small custom tool that lives inside your Figma file.

the useful bit is that it makes one-off workflow tools much cheaper to create: describe the behavior, controls, and parameters, then get a native-looking plugin panel for things like layouts, spacing, text changes, color cleanup, or data-driven visuals. classic plugins still matter for heavier external services and APIs.

🔗 https://www.figma.com/blog/agent-custom-tools-context-skills/
🔗 https://x.com/figma/status/2069824816279154840

OpenAI and Broadcom unveiled Jalapeno, OpenAI's first custom inference chip. it's designed around the "serving answers" side of AI, where ChatGPT, Codex and API requests actually run, and is meant to become a multi-generation platform deployed with data center partners from the end of 2026.

the interesting part is control of the stack: OpenAI is now designing around its own model roadmap, kernels, memory movement and networking, instead of only tuning software on existing accelerators. the performance-per-watt claim is still early though, with final numbers and a technical report still to come.

🔗 https://openai.com/index/openai-broadcom-jalapeno-inference-chip/

OpenAI published a new economic research paper on how Codex is shifting work from short chatbot exchanges to delegated agent tasks.

the most interesting bit: Codex is no longer framed as only a coding tool inside OpenAI. by June 2026, it accounted for 99.8% of weekly output tokens generated within OpenAI, and non-developer adoption was the fastest-growing segment. the caveat is important too: OpenAI is a low-friction frontier environment, so this is more a preview than a normal-company baseline.

🔗 https://openai.com/index/how-agents-are-transforming-work/

OpenAI is previewing GPT-5.6 Sol, its new flagship model, but only with a limited group first. the important part is the release process: OpenAI says Sol is stronger at coding, cybersecurity, biology, and long-horizon agent tasks, so access is being staged while safety review catches up.

that is probably the shape of frontier model launches now: better tools for defenders and builders, but slower, monitored rollouts when the same capability could help attackers too. broader access is planned after the preview, but for now this is less a normal product launch and more a controlled deployment test.

🔗 https://openai.com/index/previewing-gpt-5-6-sol/

Dilum Sanjaya demoed a practical version of vibe coding for hardware ideas: generate a 3d gadget prototype in code, test the screen/buttons/interactions, then reuse parts of the logic when moving toward a real Raspberry Pi build.

the useful bit is the process, not the polish. he starts rough, refines the shape and materials later, adds sliders for tiny tweaks, and treats the generated model as a bridge between UI experiments and physical devices.

might be a bit slow on updates but they’re still coming!!!! (On a flight) ✈️

🔗 https://x.com/DilumSanjaya/status/2070912580122755484

Meta shared Brain2Qwerty v2, a non-invasive brain-to-text research system that decodes typed sentences from MEG brain recordings without a surgical implant. the jump is in words, not just characters: Meta says it trained on about 22k sentences from 9 volunteers and reached 61% average word accuracy, with its best participant at 78%.

important caveat: this is still lab research with bulky MEG hardware, not a consumer brain keyboard. but Meta is releasing the training code for v1 and v2, plus the v1 dataset, so other labs can poke at the path from brain signals to usable communication.

🔗 https://ai.meta.com/blog/brain2qwerty-brain-ai-human-communication/

Railway made a tiny fake meditation retreat for burnt-out developers, complete with a 45-second film, “deploy code / touch grass,” and testimonials from engineers who have apparently deshrimped.

it’s marketing, but the joke lands because every cloud/devtools company is now selling speed while developers are quietly drowning in deploy anxiety, vibe-coded messes, and friday prod changes. Railway’s angle: shipping software should feel calmer, not more chaotic.

🔗 https://railway.com/peace

OpenClaw is now on iOS and Android. the useful bit is that its agents are no longer tied to a desktop or chat setup: you can run tasks, handle replies, and review approvals from your phone while the gateway stays under your control.

the iOS listing describes it as a local-first assistant paired to a private OpenClaw Gateway, with chat, Talk mode, approvals, sharing, and optional phone permissions. Android is live too, so mobile is now part of the core OpenClaw workflow.

🔗 https://x.com/openclaw/status/2071688039114342592

Extend is building document processing infrastructure for the part of AI agents that usually breaks first: messy PDFs, scans, tables, forms, handwriting, and document bundles.

the useful bit is that it's not just OCR. Extend gives teams parsing, extraction, classification, splitting, workflows, and evals, with Parse 2.0 turning complex documents into clean, agent-ready markdown instead of another pile of brittle text.

🔗 https://www.extend.ai/

Google is opening up more of its generative media stack to developers. Nano Banana 2 Lite is now generally available as its fastest, most cost-efficient Gemini image model, while Gemini Omni Flash is in public preview for video generation and conversational editing in Google AI Studio and the Gemini API.

the useful bit is the workflow: generate an image quickly, then pass it into Omni Flash to animate or keep editing through the Interactions API. it makes image-to-video feel more like a buildable product surface, not just a demo.

🔗 https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-omni-flash-nano-banana-2-lite/

Anthropic launched Claude Sonnet 5, a cheaper Sonnet model meant for agent-style work: planning, using browsers and terminals, checking its own output, and finishing multi-step coding tasks with less hand-holding.

the notable bit is positioning: Anthropic says it gets close to Opus 4.8 on agentic tasks, is now the default for Free and Pro, and is on the API at intro pricing of $2 / $10 per million input/output tokens through Aug 31.

🔗 https://www.anthropic.com/news/claude-sonnet-5

Anthropic says the Commerce Department has lifted the export controls that forced it to pull Claude Fable 5 and Mythos 5 earlier this month. access is supposed to start coming back from July 1 U.S. time, after a strange 18-day pause where a model launch became a live national-security and compliance fight.

the bigger story is that frontier model access is now getting negotiated in public with Washington, not just shipped through product pages. for builders, the question is less "is the model good?" and more "who is actually allowed to use it, and under what rules?"

🔗 https://x.com/AnthropicAI/status/2072106151890809341

Etched is out of stealth with its first inference racks in customer validation: a successful A0 tapeout, $800m raised, and $1b+ in customer contracts. it says the first racks ship this summer.

the interesting part is that this is a full systems bet, not just a chip launch. Etched is pitching rack-scale inference clusters built around low-voltage compute and shared cluster memory to make frontier model inference faster, cheaper, and less power-hungry than today's GPU setups.

🔗 https://x.com/Etched/status/2071972062202343590

oasis launched OASIS 1, a smart ring for private dictation and touch control. the pitch is less fitness tracker and more input device: whisper to write, then tap or swipe on the ring to edit, scroll, pause music, or move around apps.

the useful angle is how narrow it starts. OASIS is selling the awkward middle ground where voice is faster than typing but speaking out loud feels too public. preorders are listed at $289, with shipping around Christmas 2026.

🔗 https://x.com/oasisdevices/status/2072033581241683988

very very sorry for the lack of news!!!!!! i’ve been spending too much time in person in San Francisco, and honestly it’s been jsfhifasjkndhgfq!!!!!!!!

there’s something about being here that’s hard to get online. even at afters, every single person you meet is building something, poking at some weird idea, or asking better questions than the internet usually gives you. that kind of curiosity is contagious. lime rides all around the city, the weather helps too.

17 hours flight to go!!!!!!!!!!!!! in the meantime, come check out this event where we'll try and talk about some of the cool shit that transpired!

speak more soon!
🔗 https://luma.com/olsz24cx