Hi, it's Andreas here.
A little later than usual this week. I've been heads down on two major client projects, and things have been pretty intense.

In this issue:

Google brings Gemini 4 Argon to the Gemini app → Google AI Pro and Ultra subscribers get it now. Gemini can also run tasks while you're away, and Antigravity's coding agent now sits inside the Mac app.

Meta ships Muse Wonder and Wonder Full, its most capable models, built for agents → Wonder is available now in Meta's Model API and Muse Code; Wonder Full is for Muse Code subscribers only. Wonder will power Meta's Muse agent next.

OpenAI rolls GPT-6 and Intelligent UI out to every ChatGPT user → Answers now come with visuals and interactive elements. ChatGPT Plus, Pro, Business and Enterprise get GPT-6 Sol; Free and Go get GPT-6 Luna. Work and Codex are unchanged.

Mistral opens Large 4 in preview, with open weights due end of October
→ Mistral trains and serves it in its own data centers in Europe, with a deployment it runs under European law. It reports a top-five score on Artificial Analysis' cyber index.

Alteryx partner test: an LLM alone missed 2 breaks across 120 trades
→ With an Alteryx One workflow doing the comparison over MCP and the LLM explaining the result, the setup caught every break on 20x fewer tokens.*

Anthropic's Claude Haiku 5.5 runs about 75% cheaper than Haiku 4.5
→ Built for summaries, classification, subagents and browser use, it is the first Haiku with an effort setting. Sonnet 5.5 cache reads are now half price.

I've worked in tech for more than 12 years, and I use AI every day. You're not missing anything, and you're not "bad at AI." It's just not as good as the companies selling it say it is.

It is good at real things: writing, coding, analyzing data, making decks. I use it for all of them. But that list is a long way from what a lot of the promo videos and pitch decks promise. And when the product falls short, the message you get is that the problem is you: write better prompts, add proper context, or wait for the next model. But I believe blaming the user for what the product can't do is gaslighting.

a16z just started ranking AI apps by what U.S. consumers actually pay, based on card spending. Nearly half of U.S. consumers say they use AI. A quarter use it daily. As of August, only 4.5% had a personal paid subscription to ChatGPT, Gemini or Claude, up from 2.1% a year earlier.

The median paying user spends $25 a month, and that number has barely moved. And the big spenders, the top 1% at an average of $903 a month, look more like builders than consumers, buying automation, product-building or creative tools. That's a real market, but a pretty small one, and far from the mass market the valuations assume.

Meanwhile the money keeps pouring in, and the big AI labs' profits are nowhere near where they need to be. And when the revenue can't carry the story, the pitch has to.

Adoption isn't going well, even at tech companies. Licenses are too expensive to roll out to everyone, token costs are through the roof, compliance is a nightmare, most companies have messy data and processes, and employees are too stretched to get trained.

And people are rarely the only bottleneck. Most employees adapt faster than the approval chains, incentives and org charts around them. Organizations need more change than people do, and most are far from it. Even if the models were as good as the pitch, that gap would still be there.

Even the people who do get trained can't see where the tool stops working. Researchers call this the jagged frontier: AI is excellent at one task and quietly wrong at the next. The newest evidence comes from the labs themselves. In January, Anthropic ran a randomized trial with 52 developers learning a new coding library. With an AI assistant, they were not significantly faster, and they scored 17% lower on a quiz about the code they had just written.

I believe no tech company would hold back an amazing product it can sell. So when a lab says a model is “too dangerous” to release, I read it as part of the launch. OpenAI held back GPT-2 in February 2019, citing "concerns about malicious applications." It released the full model nine months later. I made the longer case for who benefits from these warnings in September.

The big labs are scared of cheaper competition that is catching up fast. Earlier this month, Reflection announced Beam, an open-weight model it says matches GLM-5.2 on reasoning with 3 to 4x less inference compute. Reflection still ranks Kimi K3, from a Chinese lab, ahead of Beam. The open models keep getting closer, and they cost less to run.

That's why I read the loudest "AI will end humanity" warnings with one eyebrow raised. Strict rules would hit the cheaper competitors hardest, give the big labs a reason why the promised returns are late, and let them keep shipping products that are less than amazing. I can't prove intent, and some people sounding the alarm clearly mean it. But the incentives all point the same way.

DeepLearning.AI and Qdrant teach on-device memory for AI assistants in an hour → Eight video lessons and five code examples: build an assistant that stores, recalls and forgets text, voice and image memories locally.

tldraw's guide to shipping a tool you built with AI, for non-engineers → Git, pull requests, tests, review and deployment, explained by a non-engineer at tldraw for people whose tool only works on their own laptop.

Claude Dashboards turns plain-language questions into live charts on your data → Connect BigQuery, Snowflake, Databricks or Salesforce on a paid plan (beta), ask a question, and click any number to see the query behind it.

I take on a small number of advisory engagements each quarter for AI programs that are stuck. Reply to this email with the one initiative you would want a second pair of eyes on. And if you want to build agents rather than read about them, the waitlist for the fourth Agentic AI Cohort is open, and you can start the Build Sprint today.

Reply with the one task where AI let you down this month. I read every reply.

— Andreas