brotchie @brotchie

swe @google applied ai . 🇦🇺🌁 github.com/brotchie San Francisco, CA Joined May 2008

Tweets

1K
Followers

606
Following

327
Likes

2K

brotchie @brotchie

3 hours ago

@alexandr_wang Video model human faces have successfully crossed the uncanny valley 🚀

0 0 0 10 0

View Details

brotchie @brotchie

4 hours ago

@Trowa_ @isitwokeornot The emotion expressed on the faces is unreal.

0 0 1 13 0

View Details

brotchie @brotchie

4 hours ago

@Grummz Why are there tears in my eyes, woah.

0 0 3 557 0

View Details

@tferriss @kevinrose I have this weird feeling that Kevin was AI generated, it would kinda be a random show thing to do this and then reveal later 🤣 Kevin MCP. It could be new studio / camera / lighting / editing, But Kevin-in-motion looks kinda 95% Kevin rather than 100% Kevin.

0 0 0 206 0

View Details

brotchie @brotchie

4 days ago

You don’t even need AI for this. A decent data feed of aircraft position data, arrival and departure times and you could run a MIP solver to solve for optimal gate allocation, etc. May be a layer of AI decision making on top, and you could have an AI build it in a loop, shadowing whatever system is currently used with a gradual cut over 🤷‍♂️

0 0 0 17 0

View Details

brotchie @brotchie

4 days ago

@antoniogm Depressingly accurate.

0 0 0 66 0

View Details

brotchie @brotchie

5 days ago

@creatine_cycle “Recursive self improving agent harness, you say? … Never heard of him.”

0 0 1 142 0

View Details

brotchie @brotchie

6 days ago

Opus 5.5 looping for 12 hours would beg to differ. The first pass of decomp to C isn’t great. But you get Opus to setup a TAS harness, run a 100% completion TAS, capture traces. It’s then really good at reverse engineering well named functions, structs, variable names, converts switches to if-statement all while retaining a byte identical binary 🤷‍♂️

1 0 6 1K 2

View Details

brotchie @brotchie

6 days ago

The divide comes down to whether you treat AI like a search engine or like programming. If you ask it a question, it gets it wrong, and you give up, you stay in the 75%. Levelling up happens when you want to do something genuinely hard and refuse to do it manually. You hit a wall, think "okay, let me try wiring up a 10-agent swarm for this," and 100 turns and 12 hours later you finally force it to work. That 12-hour grind is the learning curve. The 25% have been compounding those hours daily for years. Someone showing up today asking "what can AI do?" has a much more brutal upskilling curve ahead of them than they realize.

0 0 1 105 0

View Details

brotchie @brotchie

6 days ago

@four_form "Math is hard for humans, easy for AI"

0 0 0 443 0

View Details

brotchie @brotchie

6 days ago

OpenAI’s math drop nudged me toward “AI is more powerful than I thought” and massively shifted me toward “maybe mathematics isn’t intrinsically that hard: it’s just hard for humans.” The bigger update was about our limitations.

0 0 1 50 0

View Details

brotchie @brotchie

a week ago

@aquariusacquah "tell me you've never run a 12 hour agentic loop without telling me you've never run a 12 hour agentic loop"

0 0 0 93 0

View Details

brotchie @brotchie

a week ago

@Rames_Jusso @HyperFrames_ I built this without a framework, not sure I buy the premise of this post. youtu.be/BANjdYePoGI?is…

0 0 1 353 2

View Details

brotchie @brotchie

a week ago

I gave Claude 5.5 a Suno song and asked for a felt puppet music video for kids. 3 days and 840M tokens later, it explains the Big Bang, fusion, evolution, transistors, LLMs and paperclip maximizers. To toddlers. Smoosh It! 🧶

0 0 2 111 1

View Details

brotchie @brotchie

2 weeks ago

Claude 5.5 is so good at decompilation that it MUST have been a hard RL task Ant used for post-training. They've found a deep well of verifiable, human-hard engineering tasks. Take ANY binary. Set model N+1 the task to produce C code that byte-replicates the binary.

0 0 1 62 1

View Details

brotchie @brotchie

2 weeks ago

@Deltaroo3D This is the most “holy shit we’re actually in the singularity” thing I’ve yet seen. That grab, spin, release 🤌

0 0 178 10K 1

View Details

brotchie @brotchie

2 weeks ago

Input is pdfs from kitchen designer with designs, engineering plans, photos of pre-rennovated space (some with scribbled on laser measured dimensions). Photos of marble slabs. Pdfs with quotes of appliances, lighting, hardware, etc. Astra was able to essentially make 3D models in blender from most of the product spec sheets. It looked up resources on cabinet making and was able to construct all of the drawers, drawer boxes. Wood textures are generated using a shader published in a 2025 siggraph paper that simulates tree growth and then cuts slices through it. Modify it by directly talking to Astra or leaving notes on the 3D geometry and asking Astra "address the notes I just left." The entire stack was built "agentic first" so there's a fully featured CLI + MCP for modifying the plans, ensuring constraints are met, ensuring nothing broke after increasing a drawer width 1" (for example). Definitely wasn't one shotted, multiple days of work.

1 0 2 63 3

View Details

brotchie @brotchie

2 weeks ago

So the trick with Suno is what's in the []. Astra (Extra High here) adds stuff that really takes a song from "meh" to "woah." The biggest leverage thing is when asking for lyrics from Astra indicate that song structure and verse styling be communicated in square brackets. e.g. [Verse 1 | Playful, confident female lead; crisp rhythmic melody; palm-muted guitar and driving bass] Suno's built-in lyric generate a) doesn't do this, so songs lack genre-define structure and all the embellishments that make a good song sound great. Another trick is to remaster Suno outputs using Matchering against a popular song in the genre, makes a huge difference as well.