You have to use `function_call={"name": }` as a request param to force it to call a function.
@goodside
-
Atoms joke and posterior distribution sampling
By
–
With temp=2:
> Why don't scientists trust atoms? Because they make bella strengthening command hooks Pat Booking boost Heal Columbuswartupnumber204_keeper_IA= […] The joke dominates the (deterministic) posterior distribution, which you sample non-deterministically. No cache. -
Goodside: high logprobs not cached responses cause joke reuse
By
–
I don’t think this is evidence of cached responses; the logprobs may just be that high. GPT has had issues with joke reuse at least since RLHF (and likely before with FeedMe) — e.g. ChatGPT (GPT-3.5) was found to reuse the same 25 jokes for 90% responses: https://
arxiv.org/abs/2306.04563 -
OpenAI’s bug bounty excludes model output issues
By
–
Problem is you can't patch models, so it's less of a bounty thing — OpenAI's bug bounty program (emphatically) excludes all issues with model output: https://
bugcrowd.com/openai -
Overview of LLM retrieval and RAG tuning
By
–
Good intro/overview to LLM retrieval (RAG) from @c_viernickel and Bryan Zhu on Scale's blog — it's not RAG vs. fine-tuning, it's tuning for RAG; for interesting tasks you want both:
-
Researcher Access Program for deprecated models and GPT-4
By
–
There's a Researcher Access Program for deprecated models, including at least code-davinci-002. They also offer base GPT-4:
-
Mathpix & ChatGPT: Screenshots to LaTeX to Code Workflow
By
–
…or, as a few replies suggested, use Mathpix to turn screenshots into LaTeX and then ChatGPT to turn LaTeX into code:
-
Error in code generated from a mathematical formula
By
–
The formula below is from one of several recent viral threads demoing one-step math screenshot → code in ChatGPT, where the generated code is wrong. In reality, it takes several rounds of careful feedback just to correctly transcribe that formula to LaTeX:
-
Limitations of one-step math to code in ChatGPT
By
–
One-step math → code in ChatGPT is a fun demo, but not reliable. To do it for real, you likely have to:
– Transcribe the formula to LaTeX; verify; correct
– Have it explain any notation you don't know
– Generate code with verbose translation comments
– Read those comments -
Targeting LLMs via image for resume keyword stuffing
By
–
It’s been done for PDF but targeting LLMs that use conventional pdf2txt. Resume use case isn’t new but there’s more you can do with it being an image. The core trick dates back to 1990s keyword stuffing at least.