Argon, Sol & a normal week in AI (lol)
New models, agents with computers and London plans. You’re welcome.
Hey friends! 👋
There’s a particular kind of disappointment in watching AI finish a task in seconds, then spending an hour checking whether it did it properly.
A lot of what makes someone good at their job lives outside the job description. They know which document to ignore, who to ask, and why the team stopped doing something six months ago. We hand AI access to our tools, but that knowledge doesn’t automatically come with it.
Try noticing what you keep having to explain or correct this week. That’s a useful starting point for the context your AI is missing and probably a few things your next hire would appreciate knowing too.
Here's what caught my attention this week ↓
HIGHLIGHTS OF THE WEEK
Google unveils Gemini 4 Argon.
OpenAI launched Sign in with ChatGPT
Is your AI spend paying off? 7 takeaways from my latest live.
Signals
Google DeepMind unveils Gemini 4 Argon with 1M token output for complex workflows.
OpenAI launched dots - persistent AI agents powered by GPT-6 Astra, with their own cloud computer and connections to 4,000+ apps. Rolling out to Pro and Business Premium users first.
OpenAI introduced GPT-6.1 Sol for coding, computer use and professional work. API token prices are one-fifth of Astra's.
You can now use your ChatGPT subscription inside third-party products to sign in. 16 launch partners, including Notion, Vercel and Devin.
MongoDB launched Atlas Agent Engine in public preview - agent execution, persistent memory, retrieval and governance, consumption priced against existing Atlas commitments.
Cloudflare launched Threat Signals for all accounts - agentic skills that summarize security reports and extract indicators of compromise. Free tier included.
Trump signed a voluntary AI safety accord with OpenAI, Anthropic, Google, Meta, Nvidia and xAI - internal controls, external audits, board-level review. Voluntary, not enforceable regulation.
peaq integrated World ID into peaqOS - robots can request proof a person is a unique human. Seoul prescription-delivery example is a demo, not a deployed service.
Decisions API: OpenAI's new API that answers multiple-choice questions in 150ms - instead of asking an LLM to think, you give it the options and it just picks.
Muse for SMB: Meta's AI agent can now run your small business - plug in your Shopify, Slack and QuickBooks, give it a goal, and it does the work.
Inside Port’s $100M bet on software factories
I ran into Zohar, CEO of Port, outside Blue Bottle in SF, and we ended up co-working and talking about software factories: agents taking work from a ticket through to production. It sounds great until you ask who checks the work and how the agent knows how your company actually operates. That’s where our conversation got interesting: the context agents need, the decisions humans should keep, and what engineers do in between.
Who's Buying Who?
1. Bolt.new acquired Dokai. Autonomous sales agents: prospect research, account prioritization, CRM enrichment. Another AI app-builder buying work-agent capability.
2. Salesforce agreed to acquire Listen Labs. AI customer research: agents design studies, interview people, synthesize findings.
3. Replit is acquiring Atta to make it easier for anyone to explore data.
IN CASE YOU MISSED IT 👀
Here are 7 takeaways from my last live session on AI spend and ROI:
1. Automate tracking your AI spend: tools like ELI help bring costs together so you don’t have to chase invoices and update spreadsheets.
2. Map tools to workflows and owners: know what you’re paying for and who benefits.
3. Give AI context: make your policies, processes and approved examples easy to find for AI agents.
4. Match the model to the task: test cheaper models for simpler steps.
5. Define “done”: agree on a good result and how you’ll check it before automating.
6. Save repeated feedback: turn corrections into reusable instructions.
7. Count the human work: review and maintenance belong in your ROI calculation.
BASED IN LONDON?
If you're around, come hang out in person.
Catan Night for AI founders and operators. A fun night to meet fellow H2ai members, trade some sheep and find out who takes the longest road far too seriously.
AI Brew Morning: A chill chat over a coffee tasting session about what we’re building or figuring out with AI.
AI Safety Debate + Rooftop Drinks: Big AI safety questions with an all-women panel from H2ai × @ GirlCode, with rooftop drinks to keep the discussions going.
AI Startup Roast Show: 8 startups, 3 judges, and some friendly fire. Pick your favourite demo from the safety of the crowd, or if you think you have what it takes, dm me to apply to be one of our 8 demos.
H2AI community went from 0 to 101+ members in 12 hours
From Uber, AWS, McKinsey and more. H2AI is a private community for ops, IT leaders and founders making AI actually work at work. Insider insights, practical workflows and socials.
WHAT'S IN THEIR STACK?
Alex Roggero, CMO & Co-Founder at Crono. See Alex's full stack →
Want visibility on what's in yours? Map it with ELI.work
Stay curious.
Ghita