Alan

[Alan @alanpog]

@OfficialLoganK From https://t.co/GuMb9jGyTC

[Alan @alanpog]

@OfficialLoganK Why this?: “The provided URIs must be directly accessible by the end users and must not be queried programmatically through automated means. If automated access is detected, the grounded answer generation service might stop providing the redirection URIs.”

[Alan @alanpog]

@iavins YMMV. From the linked blog post:

```
(…)
- Completely removing database writes from the clients
Which will work for approximately 0% of everybody else, but we don't care about them.
```

[Alan @alanpog]

@buzzedison @aarongrider The post says https://t.co/dvjv6SMwSU, the website not the app. RN Web is a thing.

[Alan @alanpog]

@danshipper Too expensive for many use cases: $200/1M output tokens, $100/1M input tokens - for the audio portion (chat history should be priced as text and much cheaper, but still). https://t.co/KriSNAUjoF

[Alan @alanpog]

@karpathy What about models like o1, where RL on CoT, and thus language, are an essential part of what makes them work?

[Alan @alanpog]

@danshipper @every Did you use o1-preview or mini? Can you share the prompt-enhancing prompt?

[Alan @alanpog]

@OpenAIDevs When should we expect a newer cutoff date for o1-mini (Oct 2023 is kind of old for some fast changing dev libs/frameworks)?

[Alan @alanpog]

@SRinnebach @OpenAIDevs It’s already fully available in Cursor if a Pro account and usage based pricing enabled. See https://t.co/yolbq9Wuju

[Alan @alanpog]

@codecompanionai @cursor_ai @openrouter Found it. There's a field to override the openai base url.

[Alan @alanpog]

@codecompanionai @cursor_ai @openrouter How do you add openrouter models in Cursor?

[Alan @alanpog]

Can’t we have both, i.e., markdown level performance with JSON-strict outputs using a simple 2-step process (and ~10% cost increase and ~2x time):
1. Get markdown response from your favorite model (let’s say Sonnet 3.5)
2. Ask 4o-mini with json:strict to rewrite the first LLM result in your json schema

This assumes only that 4o-mini is good enough to correctly rewrite the markdown output as JSON. Which in my experience is a no-brainer assumption for most use-cases (certainly for the one in your blog post).

[Alan @alanpog]

@OfficialLoganK Gemini Live API soon?

[Alan @alanpog]

@OpenAIDevs Is there support to FT vision/images, not mentioned in the guide.

[Alan @alanpog]

@jkronand And there’s a Python version too: https://t.co/lZ0qCFqo42

[Alan @alanpog]

@VictorTaelin Why did you use a nonzero temperature?

[Alan @alanpog]

@johnjnay @MSFTResearch I hope @OpenAI allow logprobs access for GPT-4.

[Alan @alanpog]

@emollick The whole discussion on calibration seems to imply that MS used a non public chat API, with access to logprobs (like the old completions API still has). Logprobs are super useful, as this paper illustrates.

[Alan @alanpog]

@goodside I don't think so. Look how well it can spell even rarer words like "Scrumptious" and made up words like "nonversation" and "whataboutism".
https://t.co/BcSZ8nyBgp

[Alan @alanpog]

@goodside But it's not true that it doesn't know the sequence of letters of these words. And if you ask step-by-step it does the intersection task correctly.
https://t.co/RNeiWww1pK

[Alan @alanpog]

@goodside But yes, it happens even as a standalone prompt. So likely a tokenization issue.