Promise

Working on world models in Hardware(FPGA), Robotics and Virtual Experiences

[Promise @promiseeuler]

Human life is finite, the impact of purposeful action can outlast a lifetime.
Develop a strong sense of urgency, we have a lot to achieve, people.

[Promise @promiseeuler]

I was researching how hallucinations for agents in real environment simulation can be reduced, especially for world models.

The problem majorly isn't that models hallucinate. It's because it's not grounded enough to know when they should be uncertain.

A world model shouldn't just predict the next state because it has seen similar patterns before. It should constantly reconcile what it predicts with what it actually observes, update its internal state, and be able to reject its own assumptions when new evidence contradicts them.

I also think memory plays a bigger role than we often give it credit for. A persistent representation of the environment, coupled with uncertainty estimation and continuous feedback, can significantly reduce the tendency to generate plausible but incorrect states.

[Promise @promiseeuler]

@dvorahfr They do a great job for the community.

[Promise @promiseeuler]

@bindureddy The double checking has to be there really.
I am going to do a test as well.

[Promise @promiseeuler]

Better than Composer 2.5, That's good progress.

[Cursor @cursor_ai]

We've partnered with SpaceXAI to train Grok 4.5.

It’s our most powerful model yet and the first we've built for more than software engineering.

[Promise @promiseeuler]

@theo The benchmark is becoming a performative action

[Promise @promiseeuler]

@sama Nothing comes close to voice interaction.

[Sam Altman @sama]

GPT-live (next-generation voice) launches today in ChatGPT.

it feels magical and 'real'.

i have always preferred typing to talking to an AI, now i think that's going to shift.

[Promise @promiseeuler]

@bridgemindai The race of using unrealistic real world usecase benchmark to grade Model will be the end of some AI labs.

[Promise @promiseeuler]

@TencentHunyuan One of the best feature in this is the anti hallucination feature , quite a game changer for the direction of their models.

[Promise @promiseeuler]

@iam_bee01 Great content well done bee.

[Promise @promiseeuler]

@sytaylor Those words still hunt you, it was really tragic.

[Promise @promiseeuler]

@Robo_Tuo @GeekParkHQ It takes a whole lot of industry expertise to take this move into unitree.

[Promise @promiseeuler]

@OpenAI I'll be there

[Promise @promiseeuler]

@synthwavedd They are doing damage control.
Why was that said the moment gpt sol model was announced to be released

[Promise @promiseeuler]

@bnivanovX @0xSero Thank you Bobby, I’ll check it out..

[Promise @promiseeuler]

@sunnyboicrypto 🤝🤝

[Promise @promiseeuler]

One of the things I enjoy most while working on world models and researching better ways for robots to understand the real world is talking with people who are building these systems.

Hy3 from Tencent isn't a world model, but I found it interesting because of what the team chose to improve. Instead of focusing only on benchmark scores, they spent time making the model more reliable in real tasks, especially during long interactions where models often lose context or make mistakes.

That shift says a lot about where AI is heading, and it's one of the reasons I wanted to break down the research in this video.

[Promise @promiseeuler]

@ravikiran_dev7 This is largely true, data aggregating is pretty bad for some specific usecases

[Promise @promiseeuler]

@alohaproxy Corrosion is the number one reason.
It will cost a whole lot more to build.

[Promise @promiseeuler]

You can check out hy3 here.

[Tencent Hy @TencentHunyuan]

🚀Hy3 is here.

295B MoE. Best in its size class. Rivals trillion-scale flagships.
Reliable and affordable for most agentic usecases.
Apache 2.0. Friendly for commercial use.
FREE API for 2 weeks → https://t.co/EyURKwTdgi

🤗 https://t.co/twqJpqb2SL
📖 https://t.co/4uEkIU1cW4