Hello everyone,
Do you recall the post I shared about all the ‘forgotten devices’ from the past? You know, nostalgic gadgets like the Tamagotchi, Furby, Walkman, and the original PlayStation? Well, I’ve finally put together something special with that idea.
Ben Tossell@bentossell
Save which you had or wanted.
Astra image-generated all the devices + built the site.
Inspired by this (I had compiled a collection of devices a few weeks ago and was uncertain about what to do with it, so this sparked my creativity.)
Cole @colderoshay
The life of a button… 🔊
12:56 PM · Sep 22, 2026 · 175 Views
I often observe people creating impressive interactions or experiences and I just forward them to my agent to collaborate with on ideas or concepts I want to explore. Engaging in remixing and reverse engineering is a playful way to experiment.
Alongside that, I’ve developed my token activity tracker, which details the agent applications and models I utilize on a daily basis. It includes a prompt you can copy to set up your own tracker.
Ben Tossell@bentossell
Splits by agent/model; copy instructions for your own here:
bentossell.com/token-activity/
2:50 PM · Sep 18, 2026 · 9.36K Views
19 Replies · 2 Reposts · 69 Likes
Everyone’s buzzing about Jev. To be honest, I’m feeling overwhelmed by all the applications others are creating. I haven’t had the chance to use it properly yet.
Jev is now accessible to everyone. In our last post, we discussed its launch, and since then, people have been discovering new uses for it. It stands apart from typical LLMs as a tool specifically designed for builders.
With Jev, you can input text and ask questions: Is this an ad (yes/no answers)? Which folder does this belong in (select from options)? How relevant is this result (score it)?
I’ve quickly assembled a collection showcasing what others are creating with it along with a brief overview: What is Jev? Some noteworthy implementations are:
Many demos aim to incorporate Jev with current coding agents. For instance, it’s being used for instant compaction, which looks interesting but it’s shown to be a questionable idea as it defeats the cost-saving potential from prompt caching.
Sidenote: Experiments involving compaction or shedding tool calls tend to be less effective nowadays. Codex’s implementation of compaction is so efficient that the effort to establish a custom setup doesn’t justify the minor benefits you could gain for only a few weeks.
— Keshav
The CEO of TypeSafe is also focused on coding agents, having shared his insights here.
Projects in Claude Code have recently transitioned to a single master chat (instead of a folder) where users can dump their requirements, thoughts, and tasks for Claude to spawn new threads and execute the work assigned. These threads currently run as cloud sessions, but local support is on the way. This new change is still in beta and will be available to all users soon.
Also, Claude Code will finally support AGENTS.md through their newly established feature Claude Mods.
Muse, Meta’s personal AI agent, is gaining popularity among those with access. It offers a Mac app and a developer platform to build connectors, enabling users to utilize external services through Muse. Additionally, it’s heavily influenced by OpenClaw.
Muse can shop on your behalf—although it faced issues when Amazon restricted it. Meta is working on a solution, including collaborations like this one with Shopify.
Grok 4.7 has been released. It’s considered a questionable upgrade from Grok 4.6. While it performs better than GPT-5.6-Sol and Opus 5 in benchmarks, it consumes significantly more tokens, reducing its cost-effectiveness. I experimented with it and found it lacking. I was hopeful because Grok 4.6 was quite efficient for my needs in Pi, although I seldom chose it.
Are you creating a voice agent? Agent STT by Speechmatics is powered by Linden, a new speech-to-text model specifically designed for voice agents. It charges just $0.30/hour ($0.15 after discounts), making it 25% cheaper than Deepgram Flux. You can experiment with it via Pipecat or LiveKit, or directly through Speechmatics’ API. Claim $100 credit.*
-
OpenAI has established an independent group of mathematicians to assist in sharing the proofs and findings generated by its agents.
-
Test your agent in simulations with Raindrop. (They recently secured a Series A funding)
-
Transform your best X posts into an automated blog and newsletter.
-
Learn how to deploy 2,500 PRs to production in one month.
-
Explore a free, open-source Photoshop-like image editor.
-
Take a peek inside Instinct’s memory, reverse-engineered.
-
A transcription model from SpaceX AI – Grok Voice Transcribe 2.0.
-
Learn how to build a restrained software factory.
-
Powermove – a video editor that you can manipulate by conversing with a coding agent.
-
MCP vs CLI for LLMs – are we debating the wrong points?
-
Factory is now available in Slack Code.
-
On July 25, we hacked OpenAI.
-
Agent skill to demonstrate how to create a demo video of your app, incorporating zooms and voiceovers. I utilized this for my devices tweet in the introduction.
-
Exa Snapshot – Enables agents to search previous versions of the web.
-
Epoch audited 15 AI benchmarks and identified flaws in nine.
-
Natural General Intelligence – a call to develop AI models for global safety and disaster prevention.
-
Accenture’s evaluators will work within Anthropic to ensure AI safety with equivalent access as their employees.
-
Accelerating a 4B model to 730 tokens/sec on an M5 Max. Underdog is intriguing; I’ve just been onboarded and it functions like a mini operating system: email/calendar/apps, all utilizing local models without data transmission. Plus, Sigil, the founder, will guide me through it in person next week when I visit San Francisco.
Hassan@nutlope
Here’s how I operate my software factory:
- During the week, I gather ideas and inspiration.
- On weekends, I task my agent with ranking the best concepts.
- I deploy around 5-10 parallel agents to create proofs of concept.
- I discard about 60% of the outputs, refine the top contenders, and wrap up with 2-3 solid demos.
4:01 PM · Sep 21, 2026 · 15.1K Views
40 Replies · 13 Reposts · 295 Likes
Moreover, I’ve noticed a lot of discussions surrounding Jev. I am currently feeling a bit overwhelmed by the array of creations using it, so I haven’t had the chance to fully experiment with its capabilities.
We aim to cover more topics in future releases, and it appears the excitement around these tools continues to grow. Stay tuned for more updates and breakthroughs!
