Senior Software Engineer @AWSCloud. Tech talk & coding journeys here.
Boston, MA
Joined July 2008
- Tweets7.4K
- Following1.1K
- Followers809
- Likes17.2K
Pinned Tweet
Like many of us, I started running agents on different machines I own, all connected with Tailscale. I quickly found myself wanting to run builds on other machines without having to start the project on the right machine or manage multiple clones.
I had a case where I needed Linux with KVM support, and I already had a Linux box on my tailnet. So I ended up building Errand, a CLI tool that lets you run errands on other machines you own:
errand --on linux -- make test
With Errand, I can have my agents run Nix builds on my Linux box while I’m coding on my laptop. I can also run my application’s frontend and backend in Docker on my Mac mini, with the dev server forwarded to my laptop when I’m sitting at a café, so I don’t burn through my laptop’s battery.
And all that with a very simple CLI. Let me know what you think!
github.com/lydakis/errand
A realization that keeps resurfacing is that with AI, replacing shouldn’t be the goal, it should be multiplying.
Most people haven’t caught up with how the unit of delivery has changed.
With models like Opus 5.5, you no longer need to break a feature into smaller pieces. You can ship the whole thing at once.
In other words we ought to touch grass
> bro just one more AI subscription bro. i swear this one is different bro. just let me try muse wonder full bro. i'll cancel the others next month bro.
Meta has released Muse Wonder and Muse Wonder Full, achieving Artificial Analysis Intelligence Index v4.3 scores of 52 and 56 at max reasoning effort - each a material gain over Muse Spark 1.3 (max) at 48, with a release gap of only 6 weeks
Muse Wonder is available for users on Meta’s first-party API, while Muse Wonder Full is currently only available for Muse Code users. Meta notes that each model is designed for personal and professional agent use
Congratulations to @AIatMeta, @finkd, and @alexandr_wang on the release!
Key Takeaways:
➤ Both models exhibit a material jump in overall Intelligence Index score. Muse Wonder Full (max) improves on Muse Spark 1.3 (max) across all ten Intelligence Index evaluations, gaining 8 points to 56. The largest gains were seen in agentic evaluations: Terminal-Bench 4.0 (+24 points, 33% to 57%), GDPval-AA v2.1 (+176 Elo, 1681 to 1857), and AA-Briefcase (+148 Elo, 1581 to 1729). Muse Wonder (max) gains 4 Intelligence Index points and lands at 52, with notable improvements vs. Muse Spark 1.3 (max) in Terminal-Bench 4.0 (33% to 52%) and AutomationBench-AA (58% to 68%)
➤ Muse Wonder Full (max) places in the top three on six of the ten Intelligence Index evaluations, and is particularly strong in agentic knowledge work and scientific reasoning. In agentic knowledge work, it narrowly leads on GDP.pdf (33%), ranks #2 on GDPval-AA v2.1 (1857 Elo), and #3 on both AA-Briefcase (1729 Elo) and AutomationBench-AA (71%). In scientific reasoning, it tops the CritPt leaderboard at 33% and ranks #3 on SciCode with 63%. This builds on the relative strengths of Muse Spark 1.3 (max), which also showed particular ability in these two domains
➤ Both models use fewer output tokens than Muse Spark 1.3 (max). Muse Wonder (max) and Muse Wonder Full (max) each use around 54k output tokens per Intelligence Index task, down ~10% from Muse Spark 1.3 (max, 60k). This usage is lower than similarly intelligent models Gemini 4 Argon (high, 62k) and Claude Fable 5.1 (max w/ fallback, 78k), but higher than GPT-6.1 Sol (max, 38k) and GPT-6 Astra (max, 27k). Muse Wonder Full (max) uses 16% fewer input tokens per Intelligence Index task than Muse Wonder (max) (6.44M vs 7.69M), meaning it is both more intelligent and more token-efficient
➤ Cost per Task for Muse Wonder (max) lands higher than most similarly intelligent models. Despite efficient output token use, premium API pricing ($4 per million input tokens, $20 per million output tokens) and high input token use (7.69M per task) push Muse Wonder (max)’s cost to $4.52 per Intelligence Index task. This is lower than Claude Fable 5.1 (max w/ fallback, $7.63), Claude Opus 5.5 (max w/ fallback, $5.98), and Claude Sonnet 5.5 (max w/ fallback, $5.46), but higher than GPT-6 Astra (max, $3.26), Gemini 4 Argon (high, $1.99), GPT-6.1 Sol (max, $0.72), and its predecessor Muse Spark 1.3 (max, $1.60). Per-token pricing is not currently available for Muse Wonder Full, so no Cost per Task figure can be estimated at this stage
➤ Both models gain significantly in Coding Agent performance. Muse Wonder Full achieves 68 on the Artificial Analysis Coding Agents Index at max effort in Muse Code, alongside Claude Code with Sonnet 5.5 (max, 68) and Opus 5.5 (max, 66). Muse Wonder (max) scores 63. These represent material gains over Muse Code with Muse Spark 1.3 (max), which scores 54
Other details for the two new models:
➤ Context windows: 1M tokens each, unchanged from Muse Spark 1.3
➤ Pricing: for Muse Wonder, $4 input / $0.20 cached input / $20 output per 1M tokens. Pricing per token is not yet announced for Muse Wonder Full
➤ Input modalities: text, image, video, and audio
➤ Output modalities: text
➤ Availability: Muse Wonder is available on Meta's first-party API and Muse Code at launch, while Muse Wonder Full is only available to Muse Code users
This has been really fun to be part of. My dot keeps working on improving the bound, and everyone’s learning from and building on each other’s work in the PRs
We have a lot of these auto-research projects across Github now, so I built a dashboard to track them all, including contributor stats like sports teams :)
It tracks new PRs instantly and links to X mentions where everyone seems to be collaborating live.
For example I submitted the most pull requests for the integer multiplication bound, but the data shows @dysmemic actually had the most combined time in minutes with the best bound. @che_fe, @dom_scholz, @0xdoug, @lydakis, and more contributors are included.
Conditional κ = 9.32058509374204e-4 :) just submitted a new PR
app.getsupers.com/auto-resea…
George Lydakis retweeted
Prediction: the value of all verifiable tasks in the long run is 0. There is no profit there.
Anything where the success is defined in a finite and non-ambiguous way, it’s over. Value is 0.
The profits are all in taste/judgement related tasks where some like it and some hate it. Vibes only. No rationality.
Labs are hillclimbing verifiable tasks now and it might sound impressive that OpenAI found those math proofs but soon others will follow and players like DeepSeek will drive its cost to electricity.
Ambiguity is the life line and where the money is.
- Anything that’s irrational, vibe driven and controversial with no consensus definition of good is worth something.
- If the society all agree on what good looks like for a given task with no ambiguity, it’s worth 0.
- caveat: coming up with verifiable task to do is worth a lot because the ability to come up is not a verifiable task by itself
Errand 0.8.0 is out! It now supports cloud runners, and specifically @LambdaAPI:
errand --where gpu=a10 -- python3 train.py
This leases a GPU instance on the fly and runs your code there as if it were local. It’s pretty magical.
github.com/lydakis/errand
Errand now has a snazzy new website and experimental Windows support! You can send work from Mac or Linux to run on a Windows box.
Up next will be GPU support with @LambdaAPI
engineered.ooo/errand
Been playing with Vision Pro app that lets me walk around inside a transformer's residual stream. Each cloud is 256 tokens in 64 dimensions projected down to 3D.
Seeing this inside Vision Pro and being able to walk through and move things around is magical.
Been experimenting with a lightweight agent harness and it's been so nice having errand
Terminal Bench wants x86, my mac is arm, so my agents just run:
errand --where arch=amd64 -- bash harbor-bin/run.sh
and my linux box on the tailnet runs it instead
github.com/lydakis/errand
What I expected from GPT-6 Sol was a GPT-5.4 to GPT-5.5 (or 5.5 to 5.6) treatment on 6 Astra in a smaller, more efficient package. So far, it doesn’t feel like that.
It's crazy looking back that OpenAI had a code red in December about Google and Gemini 3. Look at where we are today with Anthropic.
George Lydakis retweeted
The worst part about AI engineering is how often does everyone think their weird esoteric bullshit actually matters.
I think it's a combination of sycophancy + non-determinism + models getting better so fast that you can associate model improvements with random changes you made to your environment. The result is that everyone's brains fall out as they overthink things that don't matter.