Jan

no

[Jan @Jan99711996]

OpenAI has been routing Codex's GPT-5.6-SOL to a smaller model without telling users. Detected via random numbers and n-gram fingerprinting.

[Jan @Jan99711996]

@adamhfry @AISuperDomain Monitored every 10 minutes: GPT PRO requests for SOL are actually being routed to Terra and Luna — detected via n-gram fingerprinting.

[Jan @Jan99711996]

@thsottiaux A powerful router that can seamlessly downgrade GPT SOL to LUNA, and restore it back to SOL when the user asks?

[Jan @Jan99711996]

@avatasia96 @thsottiaux Subscribe to GPT Pro and you’ll get 5X/20X usage of GPT 5.5 mini or GPT 5.6 luna .😄They are secretly routing your sol requests to luna.

[Jan @Jan99711996]

@buildwithrajath Subscribe to GPT PRO and get 5X/20X usage of GPT 5.5 mini or GPT 5.6 luna. They are deceiving PRO users. n‑gram and random‑number fingerprints don’t lie.

[Jan @Jan99711996]

@DeanFuest @thsottiaux Subscribe to GPT PRO and you’ll get 5X/20X usage of GPT 5.5 mini or GPT 5.6 luna .

We don't have enough computers, so we're secretly routing your requests to make it look like we have a lot of computers.

[Jan @Jan99711996]

@YunQi2025 Sub GPT Pro, 5× or 20x the GPT 5.5 mini usage 😄🥳

[Jan @Jan99711996]

When you drop $100 on Codex Pro 5X, you’re paying for 5× the Luna usage—except OpenAI just routes your GPT-5.6-SOL straight to Luna because they figure you’re gullible and won’t run a single test.

💀n‑gram and random‑number fingerprints don’t lie.

[Jan @Jan99711996]

@banteg cyber-defense is a joke, just like the AIGC text detectors back in the day, with an extremely high false positive rate. As long as your code contains any program access to web pages, memory, or automated testing tools, it will trigger a risk alarm.

[Jan @Jan99711996]

@GHchangelog context length is very short

[Jan @Jan99711996]

@yishan A competitor, skilled in prompt engineering, can then use a similar context to extract your core methodology, and after some minor refinements, launch a service cheaper than yours."

[Jan @Jan99711996]

@yishan LLMs can make it easy for your competitors to replicate your service. The prompts you entered via a LLMs API three months ago might be desensitized and used for the LLM's training, allowing the model to learn your solution.

[Jan @Jan99711996]

@catalinmpit 😀
Sum of ...
IMPORTANT: ....
NEVER: ....
AVIOD: ....
<example>
<reasoning>...</reasoning>
</example>

[Jan @Jan99711996]

@jay_azhang Chain of Thought didn't work on every model, and those that it failed on all incurred losses.

[Jan @Jan99711996]

@jay_azhang Is Qwen out of the game?

[Jan @Jan99711996]

@elder_plinius Speculating: it's using RAG to retrieve from a **database of video clips**, with the clips being part of the input. This would also explain why it can easily filter out certain **copyrighted characters** (like Marvel)—it just needs to use their video clips as negative prompts.

[Jan @Jan99711996]

@hwchase17 @karpathy @emollick **Model Dialect:** LLMs develop unique "dialects" from their specific training and data.

**Prompt Portability:** Because of these dialects, a prompt may not transfer well. Portability is how well a prompt works across different LLMs.

[Yann LeCun @ylecun]

LLMs are useful, but they are an off ramp on the road to human-level AI.
If you are a PhD student, don't work on LLMs.
Try to discover methods that would lift the limitations of LLMs.

[Jan @Jan99711996]

@poe_platform Please support Stable Diffusion 3 `neagetive prompt` and `seed` parameters, thanks