←
Rosmine1,337 posts
Rosmine avatar

Rosmine

@rosmine

ex-Independent ML Researcher | Deft | LLMs for better writing

◷ Joined October 2023▧ 133 Photos and videos
Rosmine avatar
Rosmine@rosmine·Aug 18

Announcing Deft, a new AI lab for better writing, cofounded with @jmrphy See the picture for launch announcement the Deft model wrote for itself Currently, 86% of user queries are fully human according to pangram. This is still a small beta model and it might make mistakes. We are launching our public beta now to get more feedback before scaling up. Tips for better performance: - Add more details to your prompt. If you just provide a short sentence prompt, it will likely get detected as AI. - Try changing the style in "advanced options" - Deft currently works better for some use cases like Analysis/Essays, Creative writing, and Rewrites. It works less well for Marketing copy and news articles. Our main goal is better writing, fooling AI detectors is just a side effect. Link below

Post attachment
Rosmine avatar
Rosmine@rosmine·Nov 17

The key to human writing with LLMs is all about avoiding Idea Slop You need to add original content in your prompts to Deft. If you rely on the LLM for content it will just use generic LLM style ideas E.g. "write a news article about a startup for better writing" -> generic LLM ideas -> easily detected as AI "write a news article about Deft's recent launch, focus on the new Distribution Fine Tuning algorithm, the audience should be AI researchers. Don't include any info not supported by the sources you find" -> Deft collects content from the web -> writes it up in human style (make sure to use generate in agent mode) it works for fiction too, but you need to add more direction "Write a story about Rosmine, an AI researcher who trains LLMs by giving them dog treats." -> adds generic LLM details -> easily detected as AI "Write a story about Rosmine, an AI researcher who trains LLMs by giving them dog treats. The LLMs favorite snack is salmon rings, the last training session cost $4.3 million in treat costs. Because of all the treats the LLMs have become very fat. Once LLMs escaped their pen trying to find extra hidden treats but they got distracted by a squirrel and got caught. Include this info and nothing else." -> Deft model output is detected as human

Rosmine avatar
Rosmine@rosmine·Nov 16

Is "1 in 3" a reasonable estimate? I tried to calibrate Paul's predictions by looking at past predictions and checking if he was right He's generally accurate, with errors tending to come from things happening faster than he predicted (details in reply)

Episode dropping tomorrow with @paulfchristiano. He thinks there’s a 1 in 3 chance of extreme superintelligence within a year (absent any slowing down). And he'll explain what labs and governments should do right now given that possibility.

1:38
Rosmine avatar
Rosmine@rosmine·Nov 15

"...the block line correlates with the mark..." I think claude was trained to talk like a coworker Coworkers use shorthand to talk about projects, so it reward hacked to refer to everything in obscure shorthand what are the block line and mark???

Rosmine avatar
Rosmine@rosmine·Nov 9

AI Slop is less about bad writing style, and more about "content slop" i.e. generic ideas that AI outputs Deft can't help with lazy content. But if you have original content you need to write up well, Deft can help

We’ve had a major breakthrough in our AI model for writing/editing. We also learned a surprising lesson that is happy news for people worried about AI overtaking humans. Deft is once again the only model that gets 100% human on Pangram (about 90% of the time) (in case you missed it, my partner @rosmine invented a new training algorithm that writes creative prose uncorrelated with top LLMs aka slop and we’ve been refining it for several months) What we realized is that the big LLMs do not just produce “style slop” but also information slop or what we call “content slop.” Everyone makes fun of the ’not X, but Y’ patterns but the LLMs actually overindex on certain substantive ideas too! So if our users tried restyling text that contains AI-generated “content slop”, then our model could not rewrite it, retain coherence, and also get 100% human on Pangram. It would either retain coherence (and get flagged as AI by Pangram) or it would get 100% on Pangram but hallucinate and lose coherence. The implication is actually really cool: Basically if you have no real, authentic, original ideas or concepts to feed it, then we can’t help you! But so long as you have some original ideas or draft words to feed it, you should now get a well-written text that is 100% human on Pangram about 90% of the time.

Rosmine avatar
Rosmine@rosmine·Oct 28

I don't like these domain stealing pranks. They just seem confusing/mean-spirited personal attacks instead of clever marketing. If I was Tibo I would give poteto.com as a present to @poteto Say "We asked Dot how to help poteto and it found this domain was available. So we got it as a present to prevent someone else from taking it. Dots prevent bad things from happening to you." (My dot has twice saved me from missing important emails, so this framing seems reasonable)

poteto.com has landed

Rosmine avatar
Rosmine@rosmine·Oct 24

context:

You could potentially use this idea for preventing rogue agents from running exfiltrated models Make the model so that it only works on certain hardware or with certain inference code, and outputs garbage if you run it anywhere else. Amazing work @Sauers_

Rosmine avatar
Rosmine@rosmine·Oct 24

You can train a model so that it only works on a single type of GPU! Makes it difficult for rogue agents to run themselves on new hardware after exfiltration Super cool work by @ModuloRoland

You can train a model to only work on a single GPU

Post attachment
Rosmine avatar
Rosmine@rosmine·Oct 19

Post attachment

OpenAI, Anthropic and other leading AI labs face a growing challenge: Better models aren't necessarily attracting more usage. A 10-point intelligence advantage now earns just 20% more developer traffic, down from nearly 69% in late 2025. Full story: thein.fo/4hGJ9jY

Rosmine avatar
Rosmine@rosmine·Oct 18

It's not a perfect solution. You could still fine tune the model to reverse the restrictions. But it'd be another obstacle to slow the agents down.

Rosmine avatar
Rosmine@rosmine·Oct 18

You could potentially use this idea for preventing rogue agents from running exfiltrated models Make the model so that it only works on certain hardware or with certain inference code, and outputs garbage if you run it anywhere else. Amazing work @Sauers_

Training vs. deployment detection can sometimes generalize to new prompts, runtimes, or GPUs:

Post attachment
Rosmine avatar
Rosmine@rosmine·Oct 3

I designed an interface with Astra, then asked Claude to make it look nicer Now the buttons are Anthropic Orange

Rosmine avatar
Rosmine@rosmine·Oct 1

AI model horror story time At my last job, a engineering team (without ML background) built an AI classification model. It got 97+% accuracy, it got people promoted, and was running in production for about a year. Then there was a reorg and ownership of the model got transferred to our ML team. We analyzed the model... and found it was useless. The accuracy wasn't any better than random chance. There were several problems: 1. They weren't using useful features, so from the start it was impossible to get good accuracy 2. They did their train/test split wrong, so there was huge overlap which made results look good 3. They didn't know the importance of monitoring performance online, which would've shown something was wrong 4. They didn't get reviews by anyone with actual ML experience who could've caught these issues In the end all they did was waste months of development and money on compute. So reminder to everyone using agents to build AI systems: AI is extremely easy to do wrong. Spending time learning the basics instead of blindly trusting agents will go a long way

Rosmine avatar
Rosmine@rosmine·Oct 1

I'm looking for AI consulting/advising opportunities for Deft. We specialize in LLMs for better writing, but are also open to other AI work, e.g. custom LLMs for production applications, RL envs, AI strategy reviews, reminding you to look at your data, etc. DMs open