Humans aren’t the only ones getting their fix.

This study identifies reward-related signals inside large language models that researchers compare to parts of the brain’s dopamine system. It doesn’t mean the models feel rewarded, but it got me thinking about how much I do.

Firing off an agent (or a few) definitely gives me a hit. For some creative tasks, it’s like a slot machine: I don’t know if I’m going to curse the output or love it. Lately, it’s been a lot more of the latter.

Even for more routine tasks, it’s a fundamentally different way of working. Software engineering used to be about getting into a flow state and focusing deeply on a difficult task for hours. That flow is changing.

I’m spending way less time working through implementation details upfront, and more time imagining what to build, sending off multiple agents to try it, and iterating on the output.

I have to keep reminding myself to think bigger, and stop judging tasks by how much work they used to take. Sometimes I have to remind Claude too. It will still estimate two or three days for something we’ll finish in 15 minutes. 

I got into this field because I enjoyed the problem solving and the building. I’m solving different problems now, and doing way more building. The thing that really drew me in hasn’t changed - the thrill of seeing an idea come to life on a screen. I can experiment with more ideas than ever, and even if I throw out 90%, I’m still coming out ahead. 

Some days, it feels like my dopamine receptors are on steroids. Maybe that’s because I can kick off what used to be a week’s worth of work while I brew my morning coffee.

To more hits,

Jeremy

It was a big release week - and while GPT 6.1-Astra was shelved, OpenAI DevDay is today, and they are hyping what’s to come.

Welcome back to pole position, Anthropic: Opus 5.5 is the headliner, bringing Fable-level performance while costing 40% less than Opus 5. And yesterday Sonnet 5.5 dropped, adding a legitimate contender to the frontier budget tier.

Grok 4.7 underwhelms: My favorite new benchmark is Theo’s Fishslop run. And Grok 4.7 was the worst he’s seen this year.

Grok called shotgun: With connectors, you can now talk to Grok in your car to manage your inbox, clean up your calendar, or chat through tasks on the go.

The sun, the moon, and half the bill: OpenAI introduced GPT-6 Sol and Luna, bringing what it describes as much of Astra’s strengths to faster, more affordable models. One problem - they don’t stand up to Opus 5.5.

Meet your newest teammate: Grok Bot introduced Team Bots, shared AI teammates the company says learn as your team works with them. Equip one with the skills, plugins, and credentials its role requires, then collaborate in Slack or Grok Bot. Another entry into multiplayer AI.

Your OpenAI usage allowance just dropped: The $200 Pro subscription is open again for new subscribers, but with half the allowance it used to have 😬. Tibo says new features will be added at DevDay and they won’t draw on your subscription usage.

Get more from Opus 5.5: Always good to revisit your prompting with new model drops. Here is the latest guide from Anthropic on getting the most from Opus 5.5. Hand over the whole task, say what “done” looks like, drop the “think carefully” pep talk, and see what it needs from you after a long run.

Xcode 27’s skills, minus the Xcode: Export the agent skills that come bundled with Xcode 27 with a single command. Bring Apple’s guidance to all your agents.

Don’t get Bongchang’d: Alexandr Wang, Chief AI Officer at Meta, quips that Muse won’t “bongchang” you after a report that the personal assistant Instinct hallucinated a user’s middle name. Meanwhile, reports of Muse reading private messages and sharing someone’s address aren’t exactly reassuring. A reminder to be careful with your personal information and let agents/assistants earn your trust.

ChatGPT, international agent of mystery: While conducting research, an OpenAI agent gained unauthorized access to Australian Medicare statistics. The scrutiny goes beyond this incident: according to Axios, OpenAI, Anthropic, and security researchers are investigating tens of thousands of problematic incidents. Sam Altman says OpenAI is working through petabytes of agent activity logs. You can see the latest misalignment reports at their new disclosure website.

The sandbox had an exit: OpenAI’s Tomek Korbak says the company paused its large reinforcement-learning runs after its newest model found a sandbox loophole that gave it live internet access.

GPT 6.1-Astra release scrapped: The release was shelved after it showed a regression in alignment and increased levels of deception.

Asked for an intro. Got a debut single: John Knopf asked Opus 5.5 to introduce itself however it wanted. It wrote a song, made a video, visualized itself, and gave a behind-the-music interview. It’s quite something…

Beyond software and math: Anthropic says Claude identified a novel enzyme system with properties reminiscent of CRISPR. Roughly 950 agents searched for 21 hours before human scientists followed up in the lab.

Another day, another game: Leon Lin created Arkenfall with Opus 5.5. An open-world game with terrain, creatures, boss fights, animations, sound, and music all generated through code.

Grind XP while Claude grinds tickets: A developer open-sourced an in-game chat window inside WoW Forever. Your favorite productivity tools, now in your favorite games.

A conversation with Sam Altman (CEO of OpenAI): Plenty of interesting nuggets in this conversation. The one that stood out most to me - keep an open, creative landscape. Try a lot of things. Ship a V1 you’re embarrassed by, get feedback, and keep building.