Noah

Noah

@nd.mt

CSE @railway.com

47 followers6 following181 posts

Noah
@nd.mt

Terra 6.1 is now on Noahbench v2 ™ Terra is a very balanced model that is fast for its weight class and fits the role of "subagent" quite well. While its comprehension (see below) severely lacks, its execution comes close to that of a higher class model (gpt 6 sol) at a far lower cost.

Noah
@nd.mt

Terra seems to be the "entry model best approachable for developers on a budget". Terra's writing prose was small, clean diffs split across multiple commits. A bit noisy but thorough.

Noah
@nd.mt

However, it often stops once the visible tests fail. Falling into "malicious compliance" territory out of sheer incompetence.

Noah
@nd.mt

Goodbye new york. A bit stressful but beautiful time.

Noah
@nd.mt

Hello people! Hobby builds are reenabled! If you are a hobby user, you are fully able to deploy on Railway end to end. Please allow a bit of time for our systems to work through a massive backlog. If there are any failures, try a redeploy and if that doesnt work lmk.

Noah
@nd.mt

Pro user deploys are back online! You should be able to fully deploy as usual, queue might take a bit longer while we work through the backlog. Please let me know if you have any questions or issues during this process!

Noah
@nd.mt

Hello folks! We have re-enabled Enterprise user builds. This is the first step of a gradual and safe rollout of our build system. Team is working around the clock to get this out to you. Hit me with any questions

Noah
@nd.mt

Everyone is writing the TypeScript compiler in rust. I wrote the rust compiler in TypeScript. Introducing “cargovroom” a rust compiler built fully in typescript.

Noah
@nd.mt

New benchmark on noahbench! Comprehension. How well can a model understand what I asked it? This bench assembles ~20 proven asks where a model fully misunderstands what i asked it. This is intended to figure out the base “intelligence” behind a model.

Noah
@nd.mt

“I sandboxed the agent” Meanwhile, the agent:

Noah
@nd.mt

New on Noahbench! Deepseek V4.1 Flash 1017 comes in 5th. This model was cheap, slow, and pretty decent.

Noah
@nd.mt

Something I've noticed in my usage, benches, and testing is that agent swarms are not useful for most projects. It feels amazing to pilot 200 agents all working towards getting your task done but, mass parallelization doesnt equal unlimited throughput. 9/10 you dont need it.

Noah
@nd.mt

I dont quite know how, but i appear to be eligible for subscriptions now! How exciting, what to post what to post.

Noah
@nd.mt

Fun poc I was toying around with. Actions & stuff running natively on Minecraft Java. Theres still some bugs but its actually quite nice, feels super fun!

Noah
@nd.mt

Noahbench ™ V2 is in the works. Adding 8 more unique tasks specifically geared towards model understanding of the core issues at hand. The goal is that a model must understand what you want to do and avoid malicious compliance.

Noah
@nd.mt

Hello, I made Apple Music in GPUI. This is quite fast and memory efficient.

Noah
@nd.mt

I've started doing lots of "pitch me the perfect world how this feels in the end" to my agents. Since then the quality output is so much better. It just understands from the start "this is where we will be". Highly suggest

Noah
@nd.mt

NEW ON NOAHBENCH ™ Muse Wonder (not full) comes in 4th!!! Slotted in just below Opus 5.5 this is the perfect price to performance to speed ratio. 10 minutes per task is mental.

Noah
@nd.mt

Gemini 4 Argon on Noahbench ™ I am not mad, just disappointed. This model is useless when it comes to real world code. Full breakdown in replies. This however actually kinda makes sense for google. I think they have given up the programming model race and are instead aiming for knowledge

Noah
@nd.mt

Right around model launches this post gets a few hundred likes each time. Its beautiful to see.

Noah
@nd.mt

I’m so thankful to live in a world where i can pre pre-order my iPhone fold 256gb Star White. What a time to be alive.

Noah
@nd.mt

FSearch 0.2 150x faster to grep a file you just edited 74% smaller index (for me 1.5g -> 388mb) 35% more files covered 65% less daemon memory (on chromium) 32% less peak memory while indexing

Noah
@nd.mt

Raycast, using fsearch. 150x faster at finding edits. 25x lighter idle. 20% more files.

Noah
@nd.mt

Hello, I made fsearch 7x faster. Find any file, in microseconds.

Noah
@nd.mt

I went to bed an woke up to ~7k followers. Still a bit doubtful this is real but wow. This is a milestone for me, very thankful to be here and for all the lovely friends I've made. More to come!

Noah
@nd.mt

Have been toying around with a Rust based file search engine. It finds files in roughly ~5ms across the *entire disk*. Uses ~60mb of ram too!

Noah
@nd.mt

New on Noahbench™ Claude Haiku 5.5 comes in 9th! Right above GPT-6 Sol. Notably: 20m a task $1.52 per task, one of the cheapest in its weight class Very solid release from Ant! Well rounded 5.5 family!

Noah
@nd.mt

Finally got access to Rex. This thing is absolutely beautiful.

Noah
@nd.mt

With everyone combining different games and mixing/matching them i set out to try something a bit more niche. Whats the FPS ceiling on minecraft for mac? The answer? ~1024 median fps at full native 5k res on my Apple Studio Display.

Noah
@nd.mt

I compared this to 20 of the top Minecraft performance mods and my singular perf mod is 8x faster at 0 quality cuts. Default settings.

Noah
@nd.mt

Jace and I made a new app to handle audio processing, routing, and control on MacOS. This app uses 18mb of ram, 0.5% cpu, and barely any impact on energy drain. By definition this is the most efficient and performant audio control app to exist for MacOS.

Noah
@nd.mt

New on Noahbench ™ GPT 6.1 Sol comes in 1st on the benches. HOWEVER, i will not be switching to this even though its high on benches. Heres why: Theres a concept I've been toying with behind "talent" and "smart". Its an odd one but it can be simply understood as "talent is the ability to

Noah
@nd.mt

mechanically perform", "smart is the ability to mechanically perform and understand *why* something is being done". I think that 6.1 sol is a wonderful refinement on top of 6, however it has not increased its base smarts.

Noah
@nd.mt

Talented models are the absolute perfect workhorses though, these models are the ones who are cheap to run, can do something when the thinking is done for them, and accomplish the goals. I am pleased with the release but will not be defaulting to it. Hats off openai!

Noah
@nd.mt

My thoughts on dev day (all speculations): Enterprise readiness is a primary focus. Ultrafast is a prime example of this, you are most likely not the target for it.

Noah
@nd.mt

Reminds me of the iphone air, its quite cool but its more of a proof "we can do thin". I expect to see future iterations of this actually be used. Dots are interesting, its all converging people realize that the ideal form factor is a personal agent. I have yet to use the implementation though so

Noah
@nd.mt

we shall see how its been pulled off. Overall i'm thinking enterprise readiness and foot in the door for agentic was the goal. I wasnt blown away by the happenings.

Noah
@nd.mt

New on Noahbench™ Sonnet 5.5 comes in p3 sitting ~4% behind opus 5.5 in task completion. This model is a bit token hungry but overall cheaper due to per token price

Noah
@nd.mt

Claude Sonnet 5.5 beats... Opus 5.5? At agentic coding?? I did not see this coming at all...