Hi folks,

Do you remember the site I one-shot with all the ‘forgotten devices’ of the past? Things like the tamogotchi, furby, walkman, og playstation, etc. Well I finally did something with it.

50 years of devices. save which you had or wanted Astra image-gen'd all the devices + built the site bentossell.com/devices/ inspired by https://t.co/8Ye3cdlYgp (i'd one-shot a collection of devices a few weeks ago and had no idea what to do with it, so this sparked my

@bentossell⁴²¹

I always see people building cool interactions or experiences and just send them to my agent to link up with an idea or something I want to explore. Remixing and reverse engineering is a great way to play with things for the sake of it.

I also built my token activity tracker which shows which agent apps and models I use daily. It’s got a copyable prompt on the site, so your agent can do the same.

built myself a harness + token usage tracker splits by agent/model, copy instructions for your own bentossell.com/token-activity/

@bentossell²⁸¹

Everyone’s talking about Jev - I’m too overwhelmed with what everyone else is building with it that I’ve not used it for anything properly yet.

Headlines

Jev is now open to everyone⁴⁸⁹. We covered the launch in the last post²⁵⁰, but since then people are finding all sorts of uses for it. It’s different from usual LLMs. It is for builders, built to be used inside a tool.

You give it some text and ask questions: is this an ad (yes/no answers), which folder does this belong in (select between choices), how relevant is this result (score something)?

I quickly built a collection of things people are building with it⁴⁹⁹ and a short explainer: What is Jev⁵⁵⁵? Some worth highlighting are:

• An extension to skip sponsor segments²³⁶ on YouTube.

• Filtering negative comments¹²⁴ out of a chat in real time.

• Search Gmail by intent¹⁶¹, or filter your to-dos⁹⁷.

• A new & smarter copy-paste¹⁶⁹, drag-and-drop¹⁰² and Cmd+F¹⁰² experience.

• Controlling your Mac with Voice¹⁹⁸.

Naturally, a fair share of demos are trying to integrate Jev with modern coding agents. For example, using it for instant compaction⁷⁶. It looks cool, but it’s a terrible idea⁶⁵. It loses the cost savings from prompt caching.

Sidenote: experiments with compaction/dropping tool calls are often a loss these days. Codex’s implementation of compaction is so clean (and good) that the pain to build a custom setup is not worth the tiny gain you can squeeze out for a couple weeks.

— Keshav⁵⁶

Coding agents are also on the mind of TypeSafe’s CEO - he wrote down his thoughts here¹⁵⁵.

Projects in Claude Code¹⁶⁹ are now a single master chat (instead of a folder) where you can dump your requirements, thoughts, tasks and let Claude spawn new threads to carry out the work you give it. These threads run as cloud sessions, but local support is coming. This new change is in beta and will be live for all users soon.

Also, finally, Claude Code will now support AGENTS.md¹⁵¹ (via their new feature Claude Mods²⁵⁰).

Muse, Meta’s personal AI agent, is getting love from everyone (who has access). It has a Mac app¹¹⁵ and a developer platform to build connectors⁶², so people can use external services through Muse. And of course, it’s heavily inspired by OpenClaw⁶⁶.

Muse can shop for you - well, it could, until Amazon started blocking it⁷². Meta is working on it, with partnerships like this one with Shopify⁶³.

Grok 4.7 is out⁷¹. It’s a questionable upgrade from Grok 4.6 - It performs better than GPT-5.6-Sol and Opus 5 on benchmarks, but takes a lot more tokens to do that, which takes away from its cost efficiency. I tried it and think it’s 💩. I was excited for it, as Grok 4.6 was pretty good for me in Pi, although I rarely chose it.

Building a voice agent? Agent STT by Speechmatics is powered by Linden⁸⁷, a new speech-to-text model purpose-built for voice agents. $0.30/hour ($0.15 after discounts), 25% cheaper than Deepgram Flux. Try it through Pipecat or LiveKit, or directly via Speechmatics’ API. Claim $100 credit⁸⁷.*

My feed

• OpenAI formed an independent group of mathematicians⁹⁸ to help them share the proofs and discoveries their agents are making.

• Test your agent in simulations with Raindrop¹⁷². (They raised a Series A recently)

• Turning your best X posts into an automated blog and newsletter¹²⁷.

• How to ship 2,500 PRs to production¹⁹⁰ in a month.

• A free, open-source Photoshop-like image editor³⁰⁷.

• A look inside Instinct’s memory¹⁹², reverse-engineered.

• A transcription model from SpaceX AI - Grok Voice Transcribe 2.0⁸².

• How to build a restrained software factory¹⁸⁸.

• Powermove¹³³ - a video editor you can change by talking to a coding agent.

• MCP vs CLI for LLMs¹⁷⁴ - are we debating the wrong thing?

• Factory is now available in Slack Code⁸³.

• On July 25, we hacked OpenAI¹¹¹.

• Agent skill to make a demo video of your app¹⁹⁴, with zooms and voiceover. I used this for my devices tweet in the intro.

• Exa Snapshot⁹⁰ - Let agents search the past versions of the web.

• People don’t want to build³³¹.

• Epoch audited 15 AI benchmarks⁷⁵ and found flaws in nine.

• Natural General Intelligence⁹⁹ - a call to work on AI models for the planet to avoid natural disasters and more.

• Accenture’s evaluators will work inside Anthropic⁶⁸ to check AI safety, with access similar to its employees.

• Speeding a 4B model to 730 tokens/sec on an M5 Max¹⁶⁵. Underdog is really interesting; I just got onboarded, and it’s like a mini operating system: email/calendar/apps, etc all with local models, no data sent anywhere. Plus Sigil, the founder, is going to walk me through it in person next week when I go to SF.

Afters

Here's how I run my software factory: 1. During the week, I collect ideas + inspo. 2. On the weekend, I give the list to an agent to rank the best ones. 3. I spin up ~5-10 parallel agents to build POCs. 4. I kill ~60%, iterate on the better ones, and end up with 2-3 solid demos.…

@nutlope¹⁵⁰

@jamesm If you ask your mom "is the sky blue" she says "yes" immediately, that's jev. If you ask your dad "why is the sky blue" he talks for ten minutes, that's an LLM.

@aarondfrancis¹⁰¹

Two days ago, GPT-6 Astra broke a yet unsolved German Army Enigma message from 1941. Amazingly Astra was able to autonomously: - Search historical archives - Compare uncertain letters - Find contextual clues - Build an Enigma simulator - Write cryptanalysis code - Run parallel …

@carterleffen¹⁰⁷

my current agents.md that made astra work well still a work in progress, i keep tweaking it try it out and see how it works for you

@argofowl¹⁹¹

all of my imessages & email now run through a single agent interface. i still text, but i deleted every email client from my phone & canceled every email subscription or app i had. gmail is basically just a database to me now. increasingly so are the rest of google’s services.…

@signulll¹¹³

You feel saas is dead when you build with Astra. Made an inbox to unify all my chats for fast triage. 4 prompts.

@ben_issen¹³²

Running Jev to see how fast it can remove negative comments from chat

@developedbyed⁸⁸

• Find me on X³⁷, Linkedin³⁵, or YouTube³⁴

• Read about me³⁶ and Ben’s Bites

• 📷 thumbnail via @keshavatearth⁵⁶

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com²⁵ or k@bensbites.com²⁷