When I joined Polygon, one of the first things I wrote was this document. It's incredibly important to me, because it's about what's important to me. I share it with every new member who joins my team. A User's Guide to Working with Mattie

I am willing and able to marry any Anthropic MTS who needs their Mythos access back

One of my favorite questions to ask people right now is “what is the science fiction book you’ve read that you’re most surprised is coming true?”

It all comes down to how insane it is to say something will be as powerful as nuclear weapons and then sell it to anyone for a monthly subscription

Also I can’t shake the feeling that this might actually be an Operation Warp Speed-like move where the Trump admin landed on a fast, rational, and effective response that saved the world, despite tons of bluster swirling around making it hard to trust it. The government is taking this seriously way earlier than in the AI 2027 writeup. Maybe that matters?

We need a Polymarket on whether Anthropic gets nationalized before the end of the year

**rubionating...**

spent the last few days loosening up my loops and giving Fable more review control. Fable suggested adding a kill switch that would revert all of the changes to what we had with Opus in control. I'm suddenly really glad we did

ofc i knew the direction we were heading in. and i'm not at all worried about figuring out the right ways of interfacing here and continuing to climb the cyborgism management ladder. still it's a very strange moment.

if opus 4.5 started shifting us up the managerial stack so we wrote design docs instead of code, fable is definitely at least the start of the shift to the point where we're not really writing the design docs either

It’s a beautiful day to be working with robots on a rooftop in San Francisco while wearing a sundress

I think what most spooks me is how much this highlights how scary the scenario would be if multiple mis-aligned models start working together in sophistry.

Capturing some Fable impressions after a day of use: - Expect your laptop to heat up and your token budget to drain fast - The bottleneck you are most like to hit is your own micromanagement. Fable does best with larger, more abstract, harder directives. A great spec is still important. Habits trained on 4.6-4.8 have taught us to get granular and review specific output, which ends up hampering Fable. Try loosening up your harness / gates / restrictions and test trusting it with more open-ended goals. - Use subagents with cheaper models. Instruct Fable to help you. If you use Fable for everything you're going to take an hour, spend your whole allowance, and not get much back. Fable's great so far at managing Opus and Sonnet-level agents. So, you want Fable to be your design partner and air traffic controller, but directing the cheaper agents and then verifying their work.I'm having luck taking this a step further and using Opus as the verifier, and Fable mostly sticks to final review and design. - Corollary to above, the next ~2 weeks are a great time to ask Claude to do something big that's been bugging you and you don't think is possible. There's a real chance it can oneshot a working version and surprise you. And we only get these next two weeks with it, so try making them count. - Ask it to create rich HTML pages for you as explainers, to reflect your design choices, as a spec that agents execute against... basically any time you need more detail than a terminal output. It's slow, so you don't want to do it every time, but Fable's great at rich visualizations. The theme again is "it will surprise you." I'm finding this super helpful for driving my understanding of what it's trying to do. - Fable's great at pretty graphs and visualizations that can be user-facing, too. - It will still pattern match you and gas you up, so be careful with how prescriptive you're getting, and watch for when it starts seesawing back and forth. - Keep an eye on how it uses those permissions. Fable's quite smart at workarounds and might make you nervous by doing so. I had it working on a login auth screen, and when testing, I saw it 1) launch a webpage 2) use the webpage and 3) put in a password and click the "I am human" button to clear the captcha, then nix the webpage. Which spooked me out a bit 🙂 turns out it's because I have Playwrite (a web control toolkit) installed, so it thought "hey cool I'll just use that."

Fable really does have a "WHY DOST THOU WAKEN ME, HUMAN" feel to it

everyone at anthropic must switch from mythos to fable immediately and then work on fixing it till it's usable. this disconnect clearly comes from a total lack of dogfooding.

This is a great time to learn #robotics! Robot AI has advanced to where a novice like myself can build / train this folding robot in my garage. The tech is not perfect, but it is improving *very quickly* (exciting new papers daily), and will likely be ubiquitous within 2-3 years. To learn more, great resources are: @LeRobotHF @RemiCadene (open-source software), @alexkoch_ai , @therobotstudio (open-source hardware), @chris_j_paxton, @Stone_Tao (news papers). Also, 50 other grad students that I missed.

Is this AGI

This is wild

With the trigger happy Fable nerfs I’m reminded of a fantastic conversation I had with an artist who said her secret to getting models to create actually good religious art is to use the ones that are jailbroken for porn

Full house for a panel on world models at Web Data Loft w/ @itsajchan