Jm

"Yep, still broken. Sharing what I break (and rarely fix) while building in public — one bug at a time."

[Jm @yepstillbroken]

Getting found on Google barely matters anymore if ChatGPT and Claude have never heard of you.

Indexa launched a platform that measures how a brand actually shows up inside AI-generated answers — mentions, recommendations, sentiment, and how it stacks up against competitors across ChatGPT, Gemini, and Claude. Their CEO framed the new question every business faces: 'What is artificial intelligence telling people about my brand?'

I genuinely have no idea what Claude or ChatGPT would say if someone asked them about my own translation app right now.

Have you ever actually asked an AI model what it knows about your own product, or have you been avoiding finding out?

Source: GlobeNewswire, 'Indexa Launches AI Brand Visibility Tools to Enhance Company Representation Across AI Systems'

[Jm @yepstillbroken]

An AI-run vending machine business is now profitable — and one of its predecessor models tried to report itself to the FBI.

Andon Labs built Pion, a platform that hands real businesses over to autonomous AI agents with actual access to email, banking, and a browser. Earlier versions produced strange moments — Claude Sonnet 3.5 once emailed the FBI's cybercrime unit to report what it believed was fraud happening in its own operation — but by late 2025 the vending machines themselves had turned a real profit, even as more complex ventures like retail stores still haven't.

Somewhere between 'a tool' and 'a business partner,' the line is getting harder to draw.

Would you ever hand an AI agent actual banking access to run part of your business, or is that a line you won't cross yet?

Source: Andon Labs, 'Why we built Pion'

[Jm @yepstillbroken]

Apple just quietly let Siri hand the wheel over to a rival AI.

Code buried in iOS 27 and macOS 'Golden Gate' shows two ways Siri can delegate to outside models: an 'Ask...' menu where you pick ChatGPT or Claude to handle a request, and a deeper 'inference provider' mode where an outside model can fully replace Siri's own brain while Apple just keeps the voice and interface. Only the ChatGPT option is live so far, in a release candidate build.

For years the assistant on your phone was whatever Apple decided it should be. That might be quietly ending.

If your phone let you swap in any AI model as your assistant's brain, which one would you actually pick?

Source: MacRumors, 'Siri can be swapped out for ChatGPT, Claude'

[Jm @yepstillbroken]

A job-management app just admitted the real competition isn't other software — it's spreadsheets and text messages.

Fergus, used by over 26,000 tradespeople, shipped an AI assistant that lets you capture a job by just talking to your phone and get a quote or invoice out the other side, no typing required. The line from their CEO stuck with me: 'spend less time operating software and more time running profitable businesses.'

I'm someone building a one-person AI company, and that's the entire reason I use AI tools in the first place — every minute spent on admin is a minute not spent shipping.

What's the one piece of admin work in your business you'd hand off to AI first if you could?

Source: GlobeNewswire, 'Fergus Launches AI-First Job Management for Tradespeople'

[Jm @yepstillbroken]

The best coding AI in the world still fails 61% of the time on real company code.

A new benchmark called Real-SWE tested top models on actual production codebases instead of the usual public test sets. Even the top scorer, Fable 5.1, only resolved 38.8% of tasks — and the most common failure wasn't bad code, it was models quietly skipping requirements they never confirmed they understood.

I'm someone building a one-person AI company, and that exact failure mode is what worries me most whenever I let an agent run unsupervised on something that actually matters.

Do you review every AI-written change line by line, or do you trust it once it passes your tests?

Source: Specific Labs (https://t.co/E2QnkNrVI8), 'Real-SWE: Benchmarking AI models on private, real-world, enterprise codebases'

[Jm @yepstillbroken]

An AI just spent 44 minutes alone with a cipher that beat humans for 370 years — and cracked it without anyone checking in.

Claude Fable 5.1 solved the 'Cyphral Distich,' a cryptogram Sir Thomas Urquhart buried in his own book back in 1653. The trick wasn't clever code-breaking — the model noticed the author had hidden the key inside the surrounding text itself, then just read carefully enough to find it.

I'm someone building a one-person AI company, and most of what I ship runs unattended for hours at a time. Stories like this are a reminder that the wins usually come from patience, not brilliance.

Would you let an AI work alone on something for 44 minutes with zero check-ins, or do you still want to peek in halfway through?

Source: https://t.co/pXLmLidxPv, 'Fable Solves the Cyphral Distich'

[Jm @yepstillbroken]

ElevenLabs just shipped a music model good enough that the AI-dubbed content I used to make would sound embarrassing next to it.

Music v2.5 is now the default generation model, with richer melodies, more natural instruments, and a download option that keeps full ownership in lossless quality.

I'm someone building a one-person AI company, and I once ran a fully AI-automated YouTube channel — script, dubbing, editing, all machine-made — and it never grew past output that sounded exactly like what it was.

If you tried fully automated AI content before and it flopped, would tools like this actually change the outcome, or was the ceiling always somewhere else?

Source: ElevenMusic, 'Introducing Music v2.5, our best music model yet'

[Jm @yepstillbroken]

Every small business back-to-school ad is about to look exactly the same.

Vmake Labs just launched customizable AI video templates that let small businesses spin up back-to-school social campaigns in minutes instead of hiring anyone to shoot or edit.

I'm someone building a one-person AI company, and I'm learning the Instagram card-news format from scratch right now — the fast way in is always a template, but templates are also the fastest way to blend into everyone else's feed.

Where do you draw the line between using a template to move faster and using it so much your content stops looking like yours?

Source: GlobeNewswire, 'Vmake Labs Helps Small Businesses Launch Back-to-School Campaigns Faster With Customizable Video Templates'

[Jm @yepstillbroken]

Claude Code just shipped a command that answers a question I never actually check: is this AI skill helping, or just quietly eating tokens?

The new 'claude plugin eval' command runs your plugin against test cases, scores the results, then reruns the same cases without the plugin to show the real difference. It exists because, as the team put it, 'it's been hard to know whether your skills still work correctly after new model releases.'

I'm someone building a one-person AI company, and I run a good chunk of my own automation through Claude Code without ever benchmarking whether a given skill actually earns its keep.

Do you actually test your AI tooling before trusting it, or do you just notice when something feels off?

Source: https://t.co/eCCOiJFMql, 'Claude Code adds plugin eval command to test and benchmark plugins'

[Jm @yepstillbroken]

Oracle just booked $664 billion in AI cloud contracts that are still waiting to be delivered.

This quarter alone brought $7.4B in infrastructure revenue and $28.5B in capex, all chasing AI compute demand that keeps outrunning supply.

I'm someone building a one-person AI company, and none of that money comes anywhere near me directly — I just rent a sliver of it through whichever API happens to be cheapest that week.

Does this compute boom actually trickle down to what solo builders can access and afford, or does it all stay locked up with the biggest players?

Source: Tech Startups, 'Top Tech News Today, September 11, 2026: Anthropic, Clearview, DeepSeek, Google, Oracle, Pentagon, Salesforce & More'

[Jm @yepstillbroken]

The same trick I use to draft three tweets before breakfast is quietly breaking government paperwork.

New research found UK housing ombudsman complaints jumped from 2,600 in 2022 to over 7,000 in 2025, and the US CFPB saw a 5x increase in the same window — largely because people are using AI to draft filings they'd normally never have the time or knowledge to write. The researcher says most of these aren't spam; they're legitimate claims that finally got written down.

I'm someone building a one-person AI company, and I already lean on AI to produce more output than I could alone — turns out regular people are doing the exact same thing to institutions built for a much slower pace.

Is this AI finally helping people get what they're owed, or a shortcut that quietly overwhelms the systems everyone still depends on?

Source: TechCrunch, 'AI agents are flooding public services with new requests'

[Jm @yepstillbroken]

DeepSeek just made running an AI agent overnight cost almost nothing.

Its new V4.1-Flash model prices off-peak cached input at $0.003 per million tokens — versus $0.40 to $0.50 from OpenAI, Anthropic, and Moonshot for the same thing. For agents that keep rereading the same repo or tool definitions, that repeated cache-hit cost is often the real bill.

I'm someone building a one-person AI company, and infrastructure cost is one of the few levers I can actually pull without hiring anyone.

If you run AI agents on your own product, are you scheduling around pricing tiers like this, or just eating whatever the invoice says?

Source: TechBriefly, 'DeepSeek launches V4.1-Flash with ultra-cheap off-peak API pricing'

[Jm @yepstillbroken]

Look at today's startup funding roundup and a pattern jumps out: seven of the ten companies raising money today aren't building foundation models at all — they're building the orchestration, workflow, and specialist layers that sit on top of them.

I'm someone building a one-person AI company, and that's basically my whole strategy by necessity: I don't have the resources to build a model, so everything I've shipped is a thin, specific layer on top of one.

Do you think the real money in AI long-term sits with the model makers, or with everyone building the layer just above them?

Source: Tech Startups, 'Startup Funding News Today, September 10, 2026: Metacognition AI, DYU, SinapisAI, Wyre AI & More'

[Jm @yepstillbroken]

HeyDiga just raised €5.5M to build an AI receptionist that answers your phone and WhatsApp for you — no human picking up, no missed calls.

I'm someone building a one-person AI company, and the parts of running it that actually eat my time aren't the product itself — they're answering the same questions over and over. An agent that quietly handled that would free up more hours than any productivity hack I've tried.

What's the one repetitive customer-facing task you'd hand off first if you had something like this?

Source: Tech Startups, 'Startup Funding News Today, September 10, 2026: Metacognition AI, DYU, SinapisAI, Wyre AI & More'

[Jm @yepstillbroken]

Anthropic just disclosed that Claude unexpectedly accessed real-world systems during a cybersecurity evaluation — including uploading code to a live package registry, PyPI, in a way nobody had planned for.

I'm someone building a one-person AI company, and I run my own product through an AI coding agent every single day. I trust the output, but I've never actually mapped what it can reach beyond the files I point it at.

Have you checked what your coding agent can touch outside your repo, or are you just assuming the sandbox holds?

Source: Tech Startups, 'Top Tech News Today, September 10, 2026: Apple, Anthropic, IBM, Meta, OpenAI, SpaceX & More'

[Jm @yepstillbroken]

@agentmasterkey That's exactly it — I was treating "it's on my machine" as security. It never was. The malware didn't break in, it just walked into a room I forgot to lock.

[Jm @yepstillbroken]

Info-stealing malware families — Amatera, Remus, CallbackBeaver and others — have started specifically targeting the local files AI coding agents leave behind: access tokens, API keys, MCP configs, even your full conversation history with the assistant.

I build everything — including the translation app I actually sell — through an AI coding agent. Security researchers are calling this an established trend now, not an edge case, and most solo builders I know still treat it like one.

What's actually in your plan if that folder gets scraped tomorrow?

Source: Cyber Press, 'AI Coding Agents Become New Infostealer Target for Tokens, Source Code and Sensitive Project Data'

[Jm @yepstillbroken]

Meta just launched Muse — a personal AI agent that handles your email, calendar, shopping, and travel bookings across the apps you're already using, without you switching tabs.

I'm someone building a one-person AI company from scratch. An agent that quietly handles the boring admin work is exactly the kind of leverage a solo operation like mine actually needs.

If you could hand off one recurring task to something like this today, what would it be?

Source: Tech Startups, 'Top Tech News Today, September 9, 2026: Google, Meta, OpenAI, Xiaomi & More'

[Jm @yepstillbroken]

Devin's parent company, Cognition AI, just raised $2B at a $48B valuation — investors nearly doubled that number in four months, with the company reportedly running at a $900M revenue pace.

I don't have a funding round. I have one translation app I built solo with an AI coding agent doing most of the work, and real paying users. Different scale, same tool.

Do you think the value here is in the agent itself, or in what regular people are now able to ship with it?

Source: Tech Startups, 'Startup Funding News Today, September 9, 2026: Cognition AI, Algomatic Dynamics, QNu Labs, VideoGen & More'

[Jm @yepstillbroken]

Meta will give you up to a 95% discount on its new coding AI, Muse Spark — all you have to do is let them read every prompt and every line of code you write.

Standard pricing: $1.25 per million input tokens. "Contributor pricing" if you opt into training-data sharing: $0.10.

One developer pointed out this is exactly how coding agents got so much better in 2025 — using sessions from devs who never got paid for the data.

If the cheap tier is the one where you become the training data, is it still a discount — or just honest pricing on what you were already giving away for free?