STUDIO 3000

This Week in the Lab · Issue #12

Milestones

The New Coke benchmark

I called my shot in Issue #2, and this week we pass the New Coke benchmark.

With considerably fewer complaints, too. The few I’ve received have mostly been about not getting it (thanks, spam filters).

Twelve issues haven’t delivered all my hopes and dreams for this newsletter. But they have delivered a newsletter.

12 notes

10,000+ words

169 stories

1,000,000 hours on X (rough estimate)

56 subscribers

12 new followers on X

And I’d like to think there was at least one above-average note.

The desire for instant gratification is real. I’d love for those subscriber and follower counts to have another zero. But 56 people choosing to hear me rant each week isn’t nothing. Thanks for coming along for this ride.

Twelve weeks ago, this was just an idea. Now it’s twelve issues - and a commitment I’ve kept.

I keep reminding myself that the work compounds. I don’t know what the numbers will look like at Issue #52 or #100. I do know I won’t get there by deciding that 56 subscribers isn’t enough to keep going.

To new milestones,

Jeremy

JUST DROPPED

The horizon keeps expanding

Claude has been doing the extra-credit assignments: Anthropic has upgraded Claude Fable and Mythos to 5.2. The company says they’re better at coding, reasoning, and vision, with output up to 30% faster than their 5.1 versions. Standard API pricing drops to $8 per million input tokens and $40 per million output tokens. Read an early-access reviewer’s impressions.

Great, now the pictures won’t sit still: ChatGPT Images 2.6 adds GIFs, so a character or product shot can become a little loop instead of another static post. OpenAI says image generation is 50% faster, with better consistency when you edit characters, products, and brand details. It’s available across ChatGPT and Codex plans, with two new models in the API.

Your bug backlog just hired a bounty hunter: GPT‑6 Cyber is OpenAI’s new model for finding security holes, checking whether they can actually be exploited, and helping develop fixes. OpenAI says it’s also better at sticking to the systems researchers give it permission to test.

The blank stare has received a software update: Emma is Nuance Labs’ new AI persona, built on Nuplex 1.0 to listen and speak at the same time. The company says she reacts with her face while you’re still talking, rather than waiting for you to finish. A limited beta API lets developers work with several personas and connect their own tools.

The AI intern has a cheaper hourly rate: Muse Ember is Meta’s small model for the work AI agents do over and over: reading documents, pulling out useful information, and handling small coding jobs. Meta says it comes close to Claude Haiku 5.5’s best Artificial Analysis score at about one-third the cost per task.

LEVEL UP

Claude would like to sleep on it: Claude can save a bad note and trip over it again later. Anthropic’s guide to agent memory explains how “dreaming” lets an agent revisit its transcripts, correct mistakes, and organize what it knows into linked wikis. If you’re building agents, the useful takeaway is to save tested lessons and current plans, keep the context loaded into every session small, and check memories against the original source. The article includes cookbooks for trying it.

Let the agent decide: Natasha saves successful agent workflows as reusable code and skills, so the next attempt needs less setup and prompting. Her testing example breaks an app into small journeys, such as creating, editing, and deleting a note. The agent chooses the relevant tests for a small change, then runs the full test suite before a bigger feature is committed.

AGENTS GONE WILD

The work is “finished”

A million-dollar bug hunt. Still cheaper than the incident meeting: OpenAI sent GPT‑6 Cyber hunting through Linux, the software underneath a huge chunk of the internet. OpenAI says its automated system developed working exploits for 948 Linux-kernel vulnerabilities that could let a local user gain administrator-level access, then generated candidate patches for review. The reported API-credit cost: $1.2 million. Eric Wallace says OpenAI is putting significant computing power toward finding and fixing open-source vulnerabilities for free.

No pages, no page numbers: University of Chicago law professor William Baude says he asked Claude to update page numbers in old constitutional law lecture notes. He came back from lunch to find a year’s worth of notes and his current syllabus deleted (everything was recovered). A small editing job turned into an unscheduled backup test.

S#*T I SAW

The new CEO is a guy called Steve. Steve is software: Skyfall’s Steve is an AI CEO that coordinates agents across support, sales, marketing, and engineering. The company says it already runs three acquired digital businesses without humans in the operating loop. Now it’s selecting five more businesses, with up to $2 million total for investment or acquisition.

Your next transcriptionist is a very good boy: Woof is Chrys Bader’s free dictation app for Mac. Press Option-Space, talk, and it pastes your words into the active text field. The site says speech recognition stays on your Mac and works offline once the model is ready. Underdog Bark 3.0 is bundled; there’s no account or subscription.

THE FOREST

See the forest, not just the trees

The real side hustle is selling you the side hustle: The perfect follow-up to last week’s note. Mark Cuban’s advice is a useful gut check for the easy-money feed: ask who profits when you buy the shortcut. Nobody who has an edge in the stock market is selling it. 

Explore Services →

Get Mirror →

Free. One email a week.

Get the next issue.

5 minutes a week to stay current on startups and AI. Get what changed, why it matters, and what's useful, along with an honest note from a very online founder.

Browse every issue or learn about The Lab.