Ben Tossell Launches Website Showcasing Forgotten Devices

On September 22, 2026, Ben Tossell released a website showcasing past "forgotten devices," bentossell.com/devices/, listing items such as tamogotchi, furby, walkman, and the original PlayStation. This tweet was posted at 12:56 PM that day and received 175 views. Four days earlier, on September 18 at 2:50 PM, he had also released another project—a token activity tracker showing which agent apps and models he uses each day; that tweet received 9.36K views, 19 replies, 2 reposts, and 69 likes.

Looking at these two websites together is more informative than looking at either one alone. They are not the same kind of product: one is a nostalgia showcase page, the other a visualization of personal usage data. But they share the same approach—which Tossell himself describes as "mashups and reverse engineering": when you see an interesting interaction or experience someone else has made, you throw it at your agent and connect it to some idea or something you want to explore. This is a way of using AI agents as an experimental bench rather than a production tool.

Regarding the devices website, Tossell's account contains three separable parts. The first is the source of inspiration: he says the website was inspired by a tweet from Cole (@colderoshay) about the life of buttons, and that he had previously generated a batch of device assets all at once but did not know how to use them. The second is the method of generation: he says Astra generated all the device images on the site and built the website. The third is the purpose: the website allows users to save devices they have owned or want.

These three parts are not equally verifiable. The website's existence, release time, and view count belong to the confirmable factual level; but "Astra generated all the images and built the website" is Tossell's own statement—the source does not explain what specific product Astra is, who provides it, nor is there independent verification. Likewise, the functional description "save devices you have owned or want" also comes from his own account. In other words, what we confirm is "he released this website and described it this way," not "this website was indeed generated end-to-end by some specific AI system."

The token tracker has a higher information density. Tossell says he built a token activity tracker showing which agent apps and models he uses each day, broken down by agent and model. More noteworthy is its distribution method: the website provides copyable prompts so that users' own agents can do the same thing.

This point creates a difference in nature between the tracker and the devices website. The devices website is a finished product; all users can do is browse and tag. The tracker, by contrast, treats "the method" itself as the deliverable—the prompts are copyable, meaning others can reproduce the same kind of tracking in their own environments. From the spread data, the tracker tweet's 9.36K views and 69 likes are also clearly higher than the devices website's 175 views, even though the two were released close together and came from the same account. Restraint is needed here: the difference in views and likes can suggest a difference in attention, but the source provides no data on users actually copying or using the prompts, so it cannot be inferred that reproduction behavior has occurred.

These two websites appeared in the same issue of the Ben's Bites article, which also relayed several product developments. Ben's Bites says Jev is now open to everyone and distinguishes it from ordinary large language models: it is designed for builders and meant to be used inside tools. Its typical use is to give it a piece of text and ask a question, such as determining whether it is an ad (yes/no), which folder it should go into (choose among options), or scoring the relevance of a result.

The Jev projects listed in that issue include: an extension that skips YouTube sponsor segments, real-time filtering of negative comments in chats, searching Gmail by intent, filtering to-do items, smarter copy-paste and drag-and-drop and Cmd+F experiences, and controlling a Mac by voice. The author also gives a clear negative judgment: integrating Jev with modern coding agents for instant compaction looks cool but is a bad idea, because it would lose the cost savings of prompt caching.

Other relayed items include: projects in Claude Code now become a single main chat rather than a folder, where users can dump requirements, ideas, and tasks, and Claude spawns new threads to execute them; these threads run as cloud sessions, with local support coming soon; Claude Code will support AGENTS.md through a new feature called Claude Mods. Meta's personal AI agent Muse has a Mac app and a developer platform for building connectors, and is heavily inspired by OpenClaw; Muse could have shopped for users, but Amazon began blocking it, and Meta is working with partners such as Shopify to resolve it. Grok 4.7 has been released, outperforming GPT-5.6-Sol and Opus 5 on benchmarks, but consuming more tokens; the author's own evaluation after trying it was very poor.

All of the above product developments come from a single source, Ben's Bites, with no official announcement links provided. This means they are at most source claims, not confirmed facts. Specifically: the source does not explain Jev's developer, the company it belongs to, or the exact meaning and timing of "open to everyone"; for the Claude Code new project structure and Claude Mods, the source only says "soon," with no exact date for release or scope of application; Muse's availability scope ("people with access") and developer platform details are unclear, and the claim that Amazon blocked shopping gives no time, scope, or official response; for Grok 4.7's improvement over 4.6 and the benchmark names and scores, the source gives no specific data.

The evaluation of Grok 4.7 especially needs to be distinguished by level. "Outperforming GPT-5.6-Sol and Opus 5 on benchmarks" is a benchmark conclusion relayed by the source, but no test names or scores are given; "consuming more tokens" is a supplement from the same source; "I think it's 💩" is the author's subjective judgment and does not constitute a performance conclusion. Lumping these three together would misread personal experience as a reproducible evaluation result.

The same issue also mixes in commercial promotion: Speechmatics' Agent STT is powered by a new speech-to-text model, Linden, priced at $0.30/hour ($0.15 after discount), 25% cheaper than Deepgram Flux. This is sponsored content; its discount conditions, how to claim the $100 credit, and the testing conditions for the comparison with Deepgram Flux are all unclear, and it should be viewed separately from editorial content.

Back to Tossell's two websites. Their real significance may lie not in the websites themselves, but in demonstrating a reproducible workflow: hand external inspiration to an agent, let the agent generate assets and assemble a finished product, then publish the method itself (the prompts). The devices website embodies "generation + assembly," while the tracker embodies "generation + distributing the method." Both point to the same judgment—for individual builders, the use of AI agents is shifting from "answering questions" to "taking on a specific small project."

But the evidence ceiling for this judgment is also clear. This article relies only on a single source, Ben's Bites, whose content is mostly the author's relay and personal opinions, lacking first-party original sources; the functions and effects of the devices website and token tracker are all Tossell's own accounts, with no independent verification. Therefore, we can confirm "he released these websites and described them this way," and we can record "what other product developments were relayed in the same issue," but we cannot assert on this basis the actual capabilities, adoption level, or commercial impact of these tools.

One observable piece of circumstantial evidence: the same issue also mentions that Epoch audited 15 AI benchmarks and found 9 of them flawed. This information is itself relayed, but it precisely illustrates why the same caution should be maintained toward all the above benchmark conclusions—when the benchmarks themselves may have problems, any relayed claim of "outperforming a certain model" is merely a claim awaiting verification.