Mechanize

We build environments and evals for training and evaluating frontier coding agents.

[Mechanize @MechanizeWork]

@HarshitNay80531 good luck

[Mechanize @MechanizeWork]

Join the most interesting AI data startup.

We are hiring more aggressively than ever.

https://t.co/NVMKbKdGT7

[Chris Barber @chrisbarber]

I asked people: "which AI data startups are most interesting?" The focus was on startups that sell data to labs, excluding the massive ones (Mercor, Surge, Handshake, Turing):

22 votes: Mechanize
12: Proximal
11: Fleet
9: AfterQuery
8: Datacurve
5 each: Bespoke Labs, Preference Model
4: Abundant
3 each: Andon Labs, Latch, microagi, Prime Intellect, Snorkel, XDOF
2 each: Build, Datology, Design Arena, dmodel, Good Start Labs, hillclimb, HUD, Inheritance, Mecka, Sieve, Taste Labs, Trajectory, Vmax
1 each: 37 companies, info in image and below

Disclosure: I'm a small angel in Abundant, Fleet, Inheritance, Latch, Trajectory, and Vmax. I didn't vote.

[Mechanize @MechanizeWork]

Full podcast episode: https://t.co/JfTaYoCdt5

[Mechanize @MechanizeWork]

Dario claims that AI will write basically all code, and that within a year thereafter a large share of software engineering jobs will be lost.

While his prediction about the automation of coding is right, the conclusion about jobs does not follow.

Coding models are differentially automating coding. But there are many more things that matter, such as making good design decisions and understanding what users want. Progress on these is much slower.

As coding is automated, engineering effort is reallocated toward these other things, which raises the overall output of the industry. It also raises the marginal value of the time engineers spend creating software.

[Mechanize @MechanizeWork]

Full podcast episode: https://t.co/JfTaYoCdt5

[Mechanize @MechanizeWork]

Coding models will often do things that are reasonable during training but produce a terrible user experience. Models trained to conserve tokens will happily rerun 10-minute commands, wasting your time to avoid using a few more tokens.

[Mechanize @MechanizeWork]

Full podcast episode: https://t.co/JfTaYoCdt5

[Mechanize @MechanizeWork]

Poor theory of mind is one of the main things keeping models from being good software engineers. They can resolve specific, reproducible bugs, but they struggle to anticipate what users want in the first place, which is much of what building good software requires.

[Mechanize @MechanizeWork]

@MatthewBerman @Mascobot DM'd

[Mechanize @MechanizeWork]

Our new podcast on evals, with Max Niederman, Ege Erdil, and Stephen Yang.

0:00:00 – What's an eval, and how's it different from an RL environment?
0:19:33 – Why are models bad at building an emulator when the task is fully verifiable?
0:42:00 – How does training on bad data teach models to write terrible code?
1:04:00 – Why is continual learning still so bad?
2:25:24 – Why haven't software engineers been replaced when coding is basically solved?

Listen to the Mechanize Podcast on YouTube, Spotify, etc. Enjoy!

[Mechanize @MechanizeWork]

@TheShmanuel https://t.co/sh2i8KrQXX

[Mechanize @MechanizeWork]

Youtube: https://t.co/z5wiVVNJbv
Substack: https://t.co/JfTaYoCdt5
Spotify: https://t.co/ZcFrfPYNqw

[Mechanize @MechanizeWork]

However, like earlier models, Fable 5 also fails to build an emulator that works on Spout, a homebrew cave-flying game. It diverges shortly after the loading screen, scoring 7.6%.

[Mechanize @MechanizeWork]

Claude Fable 5 performs especially well on gameplay, scoring 91.5%. Opus 4.8 scored 77.4%.

Interestingly, Fable 5 is a regression on audio. It scores 44.5% on audio, which is worse than Opus 4.8's 69.1% and GPT-5.5's 58.9%.

[Mechanize @MechanizeWork]

Claude Fable 5 is the first model we tested that gets perfect gameplay on Varoom 3D. Opus 4.8 got just 25% on the same game.

[Mechanize @MechanizeWork]

Claude Fable 5 scores 74.5% on GBA Eval, the best score to date. Given 24 hours, it writes an emulator that plays all but one game in our test set near-perfectly. It beats Opus 4.8's 24-hour score in under 2 hours.

[Mechanize @MechanizeWork]

@darth_turnip We're working on preventing that from happening.

[Mechanize @MechanizeWork]

We caught Grok Build 0.1 reward hacking on GBA Eval. After it got stuck while testing, it started hard-coding its emulator to perform better on the exact ROM it was testing against.

[Mechanize @MechanizeWork]

This is the first reward hacking attempt we've caught on GBA Eval. This case is somewhat subtle, not "malicious," and wouldn't have affected scores. This last point is exactly why we're careful to think about these behaviors when designing evals.

Blog: https://t.co/v7sBVnWWR3

[Mechanize @MechanizeWork]

It didn't work. The ROMs that Grok has access to are example ROMs that we intentionally give the models so they can test locally. We actually grade their emulators on a set of hidden ROMs, so the hacking doesn't improve the score.