Aditya Mehrotra

I spent the last few weeks fine-tuning an LLM with @UnslothAI The model turns vague trading strategy ideas into strict, executable backtest configs. Most LLMs are great at talking about strategies, but terrible at giving deterministic backend-valid JSON that can actually run without errors. My initial setup: • 7B param model • ~2,200 training examples • QLoRA fine-tuning • RTX 4060 Ti (16gb VRAM) I wanted to see how far I could push an open model on this. The model is far from perfect, but it's already better than the base model at generating backend-valid backtest configs Next steps: • add real production configs • increase dataset size • improve reference vs. value accuracy • compare against larger models As I train more versions, I'll track results and evals in a benchmark. More soon.

@unusual_whales It’s true btw

@RabbitsJustice @unusual_whales x.com/adityakmehrotr…

@unusual_whales If true that’s a crime against humanity

AI companies are quietly obtaining rare books, ingesting their contents, and then destroying them, even if they're exceedingly rare, per FORTUNE

Of course not. I think we need open models. But I think it’s important to point out the hypocrisy of these large tech companies. Especially bringing up their track record of opposing open source when it hurts them and embracing it when it helps them. Say in 20 years, Anthropic starts a campaign against the next tech innovation, pushing them to support open source, wouldn’t it be at least a little important to mention Anthropic’s history of being against open source? Also, do you think these companies like Microsoft are supporting open source out of the goodness of their hearts? They understand that more open models means more people will use Azure to train and fine-tune them for their own use cases (which is good)! It also lowers their reliance on expensive closed models and keeps the model market fragmented so no single lab can threaten their cloud business.

@AdityaKMehrotra @billxbf @satyanadella @JensenHuang where exactly are you trying to get with this? because Ballmer did X 20 years ago, then Anthropic has the right to be anti open weight llms?

This sounds logically embarrassing. 1. Proprietary products and supporting OSS aren’t contradictory. @satyanadella doesn’t call Linux “unsafe”. @JensenHuang doesn’t campaign against ROCm / OpenCL. Compete and coevolve. 2. LLMs are made happen by “distilling” 40 years of collective intellectual property thanks to the open culture of the Internet. Pulling up the ladder here is misanthropic for the next generation.

Maybe not @satyanadella, but Microsoft has a history of being anti-open source. Ballmer literally compared Linux to communism, and spread anti-Linux FUD through marketing campaigns like “Get the Facts”. They also lobbied the government “to stop governments from encouraging the use of open-source software” zdnet.com/article/ms-lea…

Last month: > Fable 5 gets full government ban > GPT-5.6 gets limited preview release only Today: > Opus 5 drops > Almost as good as Fable (better in some benchmarks) > Half the price > Zero restrictions They banned the “dangerous” one but gave the upgrade a free pass

@WatcherGuru Good model, according to the benchmarks x.com/claudeai/statu…

JUST IN: Anthropic officially launches new Claude Opus 5 AI model.

@Yuchenj_UW The jump from 4.8 is crazy lol

@AdityaKMehrotra this graph looks interesting.

Didn’t expect today to be Opus 5 day! Opus 5 beats Fable 5 on nearly every benchmark. It looks like a major leap in coding and agentic capabilities, with much better token efficiency. And it’s half the price of Fable 5. Can’t wait to try it.

@WatcherGuru I wonder why x.com/watcherguru/st…

The Odyssey might’ve been the greatest movie I’ve ever seen in theaters

Chipotle burrito: $15 Normal DoorDash delivery: +$10 in fees Tokens to let AI do it: +$5 Some YC backed agent does it all for $40 + monthly subscription

@Yuchenj_UW @thinkymachines Benchmarks for those wondering

We’re excited to announce that Databricks is a Day-0 launch partner of Thinking Machines Lab (@thinkymachines), bringing its first oss model, Inkling, to the Databricks platform. - the strongest US oss model - 974B total parameters, 41B active MoE - Apache 2.0 license The Thinky team is absolutely cracked. We’ve loved working with them. Come try Inkling on Databricks!

Best thing about the GPT-5.6 Sol release is that they removed the 5 hour usage limits

@goldytalks *3.2

kalshi growth team after overnight view-botting this sub 90 comment tweet to 1.3m views

@MMAUNCENSORED1 Paddy almost a billionaire from one fight! Who says the UFC doesn’t pay their fighters

UFC 329 Pay 💰 McGregor- $30 Million leads the pack. Holloway $3 Million Pimblett $850K and BSD $500K Are #UFC fighters Underpaid⁉️

@WatcherGuru There was a time x.com/stockmktnewz/s…

JUST IN: Apple sues OpenAI for allegedly stealing trade secrets.

@saylor @Strategy Just using buzz words at this point

Digital Credit is transparent because the principal market risk factor is Bitcoin, an observable, homogeneous asset. Analysts can assess BTC-related credit risk continuously, and investors can apply their own statistical models to inform valuation and trading decisions. $STRC