Ali Romman

15 y/o entrepreneur at @Alphaschool
Open-sourcing my thoughts

[Ali Romman @aliromman]

Last week, I officially sold my stake in a company for $17,000.00.

As a 15-year-old, it’s the first big step in my entrepreneurial journey, and I couldn’t be more excited.

Now, I’m on to the next chapter. Time to figure out what’s next.

[Ali Romman @aliromman]

@jeremymstamper Caught in 4K

[Ali Romman @aliromman]

@1Aakib UI is probably the thing that Sol has improved on the most compared to prior OpenAI models like 5.5

[Ali Romman @aliromman]

@FrankieIsLost Anthropic in a nutshell

[Ali Romman @aliromman]

@mil000 Notes have 0 moat

[Ali Romman @aliromman]

@gtmom Sometimes my friends will play Waymo roulette when they are leaving campus after school is over. They will order and cancel Uber rides until they get a Waymo.

[Ali Romman @aliromman]

@yacineMTB Damn what’s ur hardware?

[Ali Romman @aliromman]

@jsnnsa This is a weakness, not a strength. This means that if their text models fail, they are done for. And Sol proves that they aren’t really that far ahead, if they even are. So far their focus hasn’t resulted in any significant lead in frontier AI.

[Ali Romman @aliromman]

@rasbt True, but most of the effort options are useless. For example, Luna Max is worse in every way compared to Sol Medium.

[Ali Romman @aliromman]

@GaddipatiHarsha @AnthropicAI @OpenAI Don’t hate on virgin atlantic bro. Their software is top notch

[Ali Romman @aliromman]

GPT-5.6 Sol is the new best coding model, even better than Fable.

On DeepSWE 1.1, the most accurate coding benchmark, Sol beats Fable by 3%. It also significantly outperforms Fable on cost, token efficiency, and steps per task.

- It is 2.6x cheaper than Fable, averaging $8.39 per task vs. Fable's $21.63. This is pretty self-explanatory; no one wants to spend more money.
- It is almost 2x more token-efficient than Fable. Greater token efficiency means lower costs and higher intelligence, as a tighter context window reduces confusion.
- It takes about 30% fewer steps to complete tasks, improving context management and efficiency.

Sol stays available on subscriptions (unlike Fable) and doesn't have a panic attack when you say "biology" (unlike Fable).

Additionally, Sol helped post-train Luna, an insane feat that should be getting more attention. While UI and reasoning are more ambiguous, Sol is clearly superior in coding.

Overall, an excellent model by @OpenAI, they cooked.

[Ali Romman @aliromman]

Why DeepSWE by @datacurve is the best coding benchmark and why I picked it:

- Tasks are written from scratch, unlike other benchmarks like SWE-Bench, which use already existing PRs/commits. This means that there is no chance the model has seen the solution in pretraining (cheating).

- There are a lot of tasks, 113 across 91 repos and 5 languages. High diversity = more comprehensive.

- The tasks themselves are hard and not as dependent on prompting. The prompts are "half the length of SWE-bench Pro's, yet solutions require 5.5x more code and ~2x more output tokens."

- Verifiers are quality and handwritten, so less false negatives/positives. Also, it isn't arbitrary/subjective, which is a plus.

- More realistic, it doesn't group every single model within a 10% range like some benchmarks.

- More of a personal reason, but many respected creators (e.g, Theo) and pretty much all AI companies use and trust it, which gives it credibility.

[Ali Romman @aliromman]

@interesting_aIl Good riddance. I used that thing twice. It was horrible. Then I never touched it again.

[Ali Romman @aliromman]

@hkanji hmmm, @compliantvc i summon you

[Ali Romman @aliromman]

@birdabo easily, flash 3.5 sucks

[Ali Romman @aliromman]

@Seanfrank google crying in the corner

[Ali Romman @aliromman]

@nikitabier @hrrcnes scary

[Ali Romman @aliromman]

@mattshumer_ Wow that's crazy. Context rot?

[Ali Romman @aliromman]

Explanation on "why DeepSWE?": https://t.co/g30VTdlLS2

[Ali Romman @aliromman]

Why DeepSWE by @datacurve is the best coding benchmark and why I picked it:

- Tasks are written from scratch, unlike other benchmarks like SWE-Bench, which use already existing PRs/commits. This means that there is no chance the model has seen the solution in pretraining (cheating).

- There are a lot of tasks, 113 across 91 repos and 5 languages. High diversity = more comprehensive.

- The tasks themselves are hard and not as dependent on prompting. The prompts are "half the length of SWE-bench Pro's, yet solutions require 5.5x more code and ~2x more output tokens."

- Verifiers are quality and handwritten, so less false negatives/positives. Also, it isn't arbitrary/subjective, which is a plus.

- More realistic, it doesn't group every single model within a 10% range like some benchmarks.

- More of a personal reason, but many respected creators (e.g, Theo) and pretty much all AI companies use and trust it, which gives it credibility.

[Ali Romman @aliromman]

GPT-5.6 Sol is the new best coding model, even better than Fable.

On DeepSWE 1.1, the most accurate coding benchmark, Sol beats Fable by 3%. It also significantly outperforms Fable on cost, token efficiency, and steps per task.

- It is 2.6x cheaper than Fable, averaging $8.39 per task vs. Fable's $21.63. This is pretty self-explanatory; no one wants to spend more money.
- It is almost 2x more token-efficient than Fable. Greater token efficiency means lower costs and higher intelligence, as a tighter context window reduces confusion.
- It takes about 30% fewer steps to complete tasks, improving context management and efficiency.

Sol stays available on subscriptions (unlike Fable) and doesn't have a panic attack when you say "biology" (unlike Fable).

Additionally, Sol helped post-train Luna, an insane feat that should be getting more attention. While UI and reasoning are more ambiguous, Sol is clearly superior in coding.

Overall, an excellent model by @OpenAI, they cooked.

[Datacurve @datacurve]

GPT-5.6 tops the DeepSWE leaderboard at 73%.

Sol, Terra, and Luna results are now available.

[Ali Romman @aliromman]

@thsottiaux Don't have any usage limit resets. What's going on? I'm on the $20/m plan if that helps.

[Ali Romman @aliromman]

@kutluokan https://t.co/g30VTdlLS2

[Ali Romman @aliromman]

Why DeepSWE by @datacurve is the best coding benchmark and why I picked it:

- Tasks are written from scratch, unlike other benchmarks like SWE-Bench, which use already existing PRs/commits. This means that there is no chance the model has seen the solution in pretraining (cheating).

- There are a lot of tasks, 113 across 91 repos and 5 languages. High diversity = more comprehensive.

- The tasks themselves are hard and not as dependent on prompting. The prompts are "half the length of SWE-bench Pro's, yet solutions require 5.5x more code and ~2x more output tokens."

- Verifiers are quality and handwritten, so less false negatives/positives. Also, it isn't arbitrary/subjective, which is a plus.

- More realistic, it doesn't group every single model within a 10% range like some benchmarks.

- More of a personal reason, but many respected creators (e.g, Theo) and pretty much all AI companies use and trust it, which gives it credibility.