Hi folks,

How tf do I write an intro to the craziness that’s happened since the end of last week?!

I got access to Instinct, the personal agent all the VCs are raving about - I think it’s a bit meh? I don’t know if it’s the pro-activeness that people seem to like, but I don’t love that. Makes me feel like I’m having to do work to keep it happy, or like I’ve got a boss again - no thanks.

The first message it sent me was:

I finished working on your meeting follow-ups - reply here and I’ll send the details.

I sh*t myself a lil bit and thought, please don’t start sending emails on my behalf. And I’m pretty savvy (ish) with agents. I’ve no doubt getting these onboarding experiences for everyone is really hard, but I just don’t feel the magic yet. It also feels slow.

I’ll wait for Muse access to see what that’s like, but I don’t love the idea of giving Meta access to more info about me…they’re not the most reliable of privacy partners.

And another thing - you don’t see by default what these agents remember about you, or what context they have. I really like being able to see and edit what’s in my files to steer my agents.

Much like every app adding an AI assistant chat box in their product, I unfortunately think every AI company will start shipping their own personal agents.

Ben’s Bites is brought to you by Adobe Acrobat^135†

Acrobat’s new AI^135† turns dense reports into visual summaries, interactive reports & podcasts you can take anywhere. Sharing your work? Stylize transforms documents into polished deliverables in just a few clicks. Every answer includes clickable citations & Adobe does not train on your data. Try here^135†.

Headlines

Dario Amodei has a new essay: Pace the frontier²⁰². He says all the leading labs should slow down long enough for safety reasons. Sam Altman agrees⁸⁹, but Trump does not¹⁰⁵. He called Jensen Huang on stage at the All-In Summit: The US will not lose the AI race. David Sacks⁹⁴ (AI czar for the US govt.) adds: feel free to slow down, but no need to impose it on others. Also read:

• More AI safety takes: two ways AI goes bad¹⁸² · it’s all for the IPO¹⁴³ · pacing = losses¹⁰⁰ · fear spreads faster⁹⁹ · case for open-source⁸⁶ and quite a long summary¹³⁷ of what everyone’s actually saying.

• No IPO for OpenAI in 2026⁸⁵. Sam told Fortune’s Alyson Shontell it would be “ill-timed” given the work ahead on alignment, control and safety.

• Look ma! They are misuing claude again³⁰³ - Another Anthropic crashout.

tldraw took OpenAI up on a challenge²¹⁹. Steve (the founder of tldraw) said he could make ChatGPT’s Sketch 100x better. OpenAI’s Tibo gave him a day to prove it. The result: a whole ChatGPT-style prototype with better drawing tools built in.

this is insane - steve says he can make the sketch 100x better - tibo says do it by tomorrow - steve runs a walkthough of a chatgpt app they built with the product in it and 100x better (the products they built i in their replies - its this …w-chat-app-example.tldraw.workers.dev)

@bentossell³⁷¹

ChatGPT mini²¹⁸ - A tiny floating widget to start chats, see updates, and more. Go to Pets in your ChatGPT desktop app to switch.

Claude Code can now test whether a plugin actually helps¹⁸⁴. Run the same tasks with and without it and compare the results. Works with skills too.

Two new additions to the OpenAI API:

• GPT-Live-1¹⁵⁵ - the model behind ChatGPT’s new Voice mode. I love using it while reading books, asking about tricky terms and dictating notes. Now you can add it to your products. Here it is with Astra and a whiteboard, playing teacher¹⁷⁷.

• Agents API¹²² - OpenAI’s take on Managed Agents in the Claude API. It lets developers send any task to a Codex-like agent from their apps without worrying about configuring all the infra.

My feed

• I use Pi (the harness) a lot. Till now, you brought an API key or your Codex/Claude subs to Pi. Their new product solves that⁴⁰⁴. And I tell you what, the new DeepSeek model¹²² is really nice to work with in Pi. (I’m an investor)

• Bolt Forge¹⁸⁵ - GLM, DeepSeek and Kimi inside Bolt, free until October 14.

• ChatGPT Work has a new data agent¹⁹⁵. (Also see: summation⁹² - by Opendoor’s CTO⁵⁹)

• Should your agent app have bots or tasks¹⁴³? Maybe neither. Because both make you organise the work.

• Routines in Replit¹²⁸ - automated recurring work with AI.

• What are people building with GPT-6 Astra³³⁴?

• Assistant Benchmark¹¹⁸ tests everyday assistants like Grok bot, Muse, Instinct and more. Ratings are a work in progress.

• SF autoresearch⁹² - Run agent-driven ML experiments while you save 25% on GPUs⁴⁷ on average.

• A coding agent to improve your customer-facing agents¹¹¹.

• How Pangram detects AI writing¹⁹³ - and why a human rewrite can still get flagged.

• How to run a team of agents²⁴⁴, with agents as their managers.

• Cognition’s SWE-2¹⁰³ (based on Kimi K3) beats Grok 4.6 at half the cost.

• An interactive explainer of the AI compute stack¹⁸⁵.

• Inspo²⁶⁸ - Design inspiration from 800+ sites, available as an MCP server.

• Core Auto is hiring interns¹⁰² to be the human edge for businesses run by agents.

• Why even good AI startups end up with bad prompts¹³², and how to fix them.

• Andrew Ng’s series on the skills AI engineers need³⁰¹.

• Underdog²⁷⁷ - a personal AI that runs entirely on your device.

Afters

I must be among an extremely small group of people (n=1?) that have both 1) trained a frontier LLM and 2) designed and synthesized custom viruses in a lab with my own two hands. And I think that the takes on AI killing us all by creating dangerous viruses is total bogus.

@DavidRBellamy¹⁴¹

Prompt I use a lot: "how can I pokayoke this so this kind of error never happens again" pokayoke is using simple, low-complexity methods to make a particular kind of mistake impossible to make. classic one would be connector shape: make it impossible to insert "the wrong way".

@nateberkopec¹²²

i scraped 85 million birthdays and used astra to make a website that tells you how rare your birthday is. birthdaytraffic.com

@adriannalakatos³²⁰

idk about anyone else, but I find it really hard to dictate what I'm thinking. typing unlocks something in my brain that speaking just doesn't.

@almonk⁸²

I talked to Sid & Robert about building Claude Code: how much things have changed, how hard it's been to keep up with model capabilities and also what we miss about software engineering before AI.

@trq212⁹⁴

When you want Astra to review snapshots of your app, ask it to collage them into a contact sheet to minimize the number of image read calls it needs to do. You’ll save a ton of context usage This works for fable too!

@pvncher⁹⁰

I sat down with AI leaders from @figma, @Box, @cognition, @tryramp, BAM, and the one and only @clairevo to hear how they’re using Astra. Optimizing GPU kernels, revamping workflows with computer use, designing many consistent screens… Amazed by what they’d done in just days! 🌌

@romainhuet⁸⁴

On macOS and wanna see something interesting? Ask your agent: "look at ~/Library/Application Support/Knowledge/knowledgeC.db and tell me some interesting facts" I had no idea this stuff was all getting logged and it can infer a lot about your activity.

@cooperx86¹¹⁹

ok everyone, i took one for the team here's the data we all wanted to see - real token value of each LLM subscription, empirically measured through usage on my real subscriptions - supergrok heavy has now become the highest value at $12k worth of tokens (40x ROI) - the $200 …

@kunchenguid¹¹⁶

• Find me on X³⁸, Linkedin⁴⁰, or YouTube⁴⁰

• Read about me³⁸ and Ben’s Bites

• 📷 thumbnail via @keshavatearth⁴⁴

* sponsors who make this newsletter possible :)

Wanna partner with us for the next quarter?

Email us at shanice@bensbites.com³² or k@bensbites.com³⁰