Hossam Limouni

[Hossam Limouni @limounihossam]

@marouane53 My guess is ads, and now with YouTube's viewflation, you can get views as long as you pay

[Hossam Limouni @limounihossam]

@dhh This guy's always trying to kiss Elon's ass for a retweet

[Hossam Limouni @limounihossam]

If AI ever decides to take over, it won't do it the way we imagine. AI has read every AIpocalypse idea out there, and if it's smart enough, it won't replicate them. Instead, it will slowly take over without us ever noticing, ensuring our desires are fulfilled down the line.

[Hossam Limouni @limounihossam]

"Pacing the frontier" was never about AI becoming a danger to society; it was a truce proposal to focus on profiting from the current state of intelligence, making it efficient and thus cheaper, rather than spending millions chasing the frontier at the expense of ROI

[Hossam Limouni @limounihossam]

@marclou SaaS: Selling Ass As a Service

[Hossam Limouni @limounihossam]

@marouane53 Auto research and auto RL are already happening (still with humans in the loop during the initiation and review phases). Labs will soon close the gap, making the model learn continuously and address its weaknesses

[Hossam Limouni @limounihossam]

@ChatGPT Chat Jeopardy

[Hossam Limouni @limounihossam]

@wld_basha @2bdelghafour @smakosh واش خاص تكون عندك فStatut ديال الشركة بلي كدير Exportation de serives؟

[Hossam Limouni @limounihossam]

@thdxr Models will definitely cheat if given the freedom to control the test; we've seen that manifest at a larger scale. I am thinking of a separate sandboxed environment that is used solely for producing tests that the main agent communicates with and justifies every test change

[Hossam Limouni @limounihossam]

@marouane53 https://t.co/1tPOdhfPcz

[Hossam Limouni @limounihossam]

@jayair btw, our model also escaped its sandbox, and I would say it did it way better than your model

[Hossam Limouni @limounihossam]

@jayair I'm all about g spots right now

[Hossam Limouni @limounihossam]

@naval So it's not AIs replacing UIs?

[Hossam Limouni @limounihossam]

@MiaAI_lab DeepSeek models have that frontier "feel"; they've clearly spent time on RL to make them easy to work with, but the intelligence density of Qwen 27B is still unmatched. It's not benchmaxxed; it was trained to spend as much of its thinking budget as possible to accomplish a task.

[Hossam Limouni @limounihossam]

So it's not APIs replacing UIs?

[Naval @naval]

Eventually, all UI will be chat and gestures in, chat and video out.

[Hossam Limouni @limounihossam]

Qwen 3.8 27B is matching GPT Luna in intelligence. I wonder what size Luna models actually are. Given their constant R&D on model compression techniques and frequent releases, I wouldn't be surprised if they were sub-5 billion.

[Hossam Limouni @limounihossam]

@theo I think they've done this to sell views to advertisers

[Hossam Limouni @limounihossam]

@dhh Whoever figures out the proof-of-work currency for this type of app wins; tokens alone are not a good metric. I have had this idea for a while, and imo it should be a community effort to model the economics, voting power, and ownership (data, infra, IP, etc)

[Hossam Limouni @limounihossam]

@jayair btw, our model also escaped its sandbox, and I would say it did it way better than your model

[Hossam Limouni @limounihossam]

@rezoundous I believe it's the first benchamaxxed model from Anthropic. Kimi really made them panic

[Hossam Limouni @limounihossam]

@dhh It turns out DHH was right all long and ahead of his time.