Engineer • Building Farlix I write about Frontier, AI, Agents & SaaS Products
- Tweets 3,004
- Following 365
- Followers 533
- Likes 6,807
$50 plan is the sweet spot, and there are few reason for it.
1. It gives you little higher revenue generating users.
2. Pro plan with no 5-hour limit makes more sense than hitting a limit in just a prompt.
3. Cost is going high and $20 plan making no sense here.
4. They want to try out to see that if users are acomikg over there, nad they come it means people are willing to pay for AI benifits.
5. It cost real deal of GPUs to manage efficient usage.
Anyway, the free subsidy is coming to an end and brace for the price hike in AI. No one is ready to accept this but this is going to get real.
3
127
This is insane from the Mira Murati, former OpenAI CTO:
Current models are still turn-based. You talk, they go quiet and think, and while they’re thinking they’re effectively deaf and blind. You finish your thought, wait, then try to pick up the thread.
The next ones don’t work that way. They stay open to audio, text, and video the whole time and respond in small continuous chunks. So instead of waiting for you to finish and then catching up, they can cut in the moment you start heading the wrong way.
Prompt box → chatbot → agents → an AI that is actually working next to you in real time.
Anyone can type a prompt. Very few people are ready for something that never fully stops paying attention.
x.com/sohmehta/status/211903…
2
7
516
DDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDDD
CCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCC
since u guys asked for this
1
2
187
Dhaval retweeted
A 3D arcade racer from GPT-6.1 Terra in Codex is a clean one-model build, and shipping a track you can actually drive is what makes it stick. The Last Signal is on aicookd → aicookd.com/c/the-last-signa…
1
1
2
27
All the hype about GPT-6.2 and the Fable 5.5 is dying.
People were all excited about those models through last 2 weeks and did not get anything. All the things we saw on X were completely fake.
I worte the below post, when the hype was getting built and stated the reason and seems both reason coming out of be true.
Both the labs wants to maintain the hype and don't want to release anything.
So, probably my theory is coming to the right conclusion: we did not get anything.
That comes down to 2 things:
1. We see all on the X is fake, it just a hypt and all fake things.
2. All is true, and it seems real security concerns are going on
If none of these things are there, then it's just that they don't want to release anything and maintain the hype.
3
6
422
Dhaval retweeted
Claude Ultra has no five-hour limit for verified accounts. It still has a weekly one, and one Ultra user says it was gone in 3 days.
OpenAI's Pro plans drop the five-hour limit in Codex too.
How many days does your week last?
ALT Where the five-hour and weekly usage bars are on Claude and ChatGPT/Codex, quoted from their help pages
1
5
216
Tibo should stay in the basement for a few more days; I think if not a model, then we at least get some really good games from him to play here on X.
btw, it was a good one, and I finished it, though!
1
1
7
366
@thsottiaux what do you think, it's deal, you stay in basement and delivery a really good game on poteto.com?
1
45
Nous Research just showed how easily AI model behavior can be changed!
They took Qwen3.8-27B and performed fine-tuning on just 53 million tokens to limit responses consistent with Chinese government propaganda.
Results:
- Before fine-tuning: 22%
- After fine-tuning: less than 1%
If 53 million tokens are enough to change the model's political behavior so significantly, how much of what we consider its personality comes from fine-tuning?
x.com/NousResearch/status/21…
Introducing Arete-27B: Qwen3.8-27B, post-trained to reinforce Western values.
Chinese open models are currently the strongest available, but they ship with state-aligned behavior in the weights. We built an 803-prompt eval across seven value categories, including freedom of expression, democracy, and privacy, and measured how often a model adopts China’s party line.
After fine-tuning on roughly 53 million tokens, Arete-27B's propaganda rate fell from 22% to 1%, in line with the Western models we tested. Its Hermes Bench score shows no significant performance tradeoffs.
The values in an open model are one fine-tune away.
Our eval, weights, and training data are available now along with the full blog post: nousresearch.com/arete
1
1
6
286
