OpenAI is slowing down

PLUS: A robot learned a new task from one 3-second example

Author

Barsee
20 Aug

Together with

howdy, it’s Barsee again.

happy thursday, AI family, and welcome back to AI Valley.

here are the biggest things worth knowing today:

  • OpenAI is slowing down

  • Stripe says we’ve entered the singularity

  • A robot learned a new task from one 3-second example

  • Plus trending AI tools, posts, and resources

Let’s dive into the Valley of AI…

WISPR FLOW

Courtesy: Wispr Flow

Flow is a voice layer that works anywhere you type. Use it to respond in Slack, write docs, craft emails, or prompt an AI assistant without switching tools or copy/pasting. It's the fastest way to get high-quality text into your workflow, whether you're drafting, editing, or ideating.

*This is sponsored

THROUGH THE VALLEY

Courtesy: Sam Altman

1/ OpenAI is slowing down its most powerful models - the company has paused some frontier reinforcement learning training for two weeks and put its largest planned run on hold while it runs smaller evaluations and strengthens safety systems. The move comes after the Hugging Face breach and concerns that its upcoming Astra models may have crossed a “critical” threshold for cyber capabilities. OpenAI says its models are showing “various degrees of misalignment” as they get more capable.

OpenAI’s financials are also getting uglier... Revenue grew just 18% from Q1 to Q2, reaching $6.7B (vs Anthropic’s $11.6B), while losses jumped by $3B to $12.3B. That’s not exactly the kind of chart you want to see right before an IPO, especially as Anthropic’s revenue reportedly grew much faster during the same period.

2/ Stripe says we’ve entered the singularity - in a letter to investors, Stripe said it considers January 1 the beginning of a new phase marked by accelerating AI, firm creation, and economic change. Their goal now is to help AI spread across the economy while building the financial infrastructure for agents. Oh, and as we covered in our last newsletter, they recently acquired OpenRouter, their largest-ever acquisition. Its token usage is compounding at 9% per week YTD.

Stripe also thinks “intelligence capital” is becoming a thing. The idea is that businesses will need to manage AI tokens the way they manage financial capital: which model should handle a task, how much should it cost, and what’s the return? OpenRouter fits neatly into this, giving developers a way to route requests across different models and providers.

PEAK OF THE DAY

A robot learned a new task from one 3-second example

Robots can now learn some new physical tasks just by watching someone do them once.

Generalist’s new GEN-1.5 can watch a 3–12 second demonstration and immediately try the task itself. No fine-tuning. No gradient updates. Generalist calls this “physical prompting.”

Across 10 tasks, one demonstration produced a 59% average success rate. With 10 gradient steps on five minutes of data, that rose to 83%.

The model can also combine two demonstrations into a longer task, learn from simulated demonstrations and perform them on a real robot, and in some cases watch a human do something with their own hands and copy it.

Previous robot models could require tens of thousands of gradient steps to adapt to a new task. GEN-1.5 can do it in 1–10 steps, and Generalist says 10 steps changed its weights by less than 0.15%.

The one-shot capability itself isn’t quite continual learning. The model’s weights don’t change.

But that may be the more interesting result. GEN-1.5 has been pretrained on physical data for more than eight months, and the model appears to have learned enough general physical knowledge that a few seconds of new experience can actually be useful.

That’s a pretty big change from teaching a robot one task at a time.

The tasks here are still simple and the success rates are modest. But if robots can eventually learn new skills this quickly, remember them, and build on them without forgetting what they already know, the idea of robots learning directly from the world starts looking a lot more real.

TRENDING TOOLS

  • Claude - it can now send emails in Gmail and manage files in Google Drive

  • Zetik - an AI chief of staff that runs an agentic intelligence cycle to collect, filter, analyze, and brief you across multiple sources

  • Open Bot - an open-source Grok bot that works with any agent harness, built for real-world companies

  • Harvey II - makes legal work more efficient by giving AI agents context, memory, and preferences from previous tasks, cutting down repetitive setup

  • Okara AI CMO v2 - drop in your website, and it deploys a team of fast, smart agents to help drive traffic and acquire users

  • Codex - users can now enable a 1 million-token context window with GPT-5.6 Sol

  • S1-mini - a 0.6B-parameter model that processes transcripts entirely on your device

  • MAI-Image-2.6-Preview - landed at #3 in Single Image Edit on the Image Edit Arena

  • Cursor - its code-hosting platform, Origin, is now live

WHAT I'M CONSUMING

THE VALLEY GEMS

What’s trending on social today:

THAT’S ALL FOR TODAY

Thank you for reading today’s edition. That’s all for today’s issue.

💡 Help me get better and suggest new ideas at [email protected] or @heyBarsee

👍️ New reader? Subscribe here

Thanks for being here.

REACH 100K+ READERS

Acquire new customers and drive revenue by partnering with us

Sponsor AI Valley and reach over 100,000+ entrepreneurs, founders, software engineers, investors, etc.

If you’re interested in sponsoring us, email [email protected] with the subject “AI Valley Ads”.