OpenAI released GPT4-API: everything you need to know!

BoredGeekSociety
6 min readMar 14, 2023

The new multi-modal model, that ranks in top 10% at the bar exam, powering Bing AI!

A short post to summarize what you need to know about GPT-4, and there is a lot to cover!

As a matter of fact, today marks an exciting milestone for OpenAI with the release of their first multi-modal API, the very anticipated GPT-4!

With already several early adopters such as Morgan Stanley, Khan Academy, and the Government of Iceland!

You can find the announcement link here:

Everything you need to know! But first, If you like this topic,
please consider supporting us: 🔔 clap & follow 🔔

What do we know about GPT4? Summary

  • Summary of the model’s capabilities: It is indeed the first multimodal model released by OpenAI since it will* be able to take both text and images as an input. It remains limited to text only for its output though. It is way smarter than ChatGPT when it comes to complex reasoning. But we’ll get there in a bit.
GPT4 analyzing a meme (source)
  • API Access: there no public release yet. Access goes through a waitlist that you can find here: join the waitlist.
  • *Really multimodal? weeelll eventually yes! Handling text and images as input. But for the early access to those in the waitlist, GPT4 is still a text only GPT4! “Once you have access, you can make text-only requests to the gpt-4 model (image inputs are still in limited alpha)” (source).
    This is aligned with our previous predictions!
  • GPT4 access available with ChatGPT plus: “ChatGPT Plus subscribers will get GPT-4 access on chat.openai.com with a usage cap.”
  • API prices: We see a sharp increase in the API price compared to ChatGPT. Depending on the GPT4 model used, token prices vary. For context, ChatGPT API is priced at $0.002/1k today. New prices are at least x15 higher depending on the model used.
    GPT4–8k token context: “Pricing is $0.03 per 1k prompt tokens and $0.06 per 1k completion tokens.”
    GPT4–32k token context: “Pricing is $0.06 per 1K prompt tokens and $0.12 per 1k completion tokens.”
  • Larger context means more complex applications: A 32k tokens context means about 25 000 words or 50 pages! This is huge! It means it can process full reports, short books or long chapters, taxation code, and so on, enabling scenarios that were impossible with ChatGPT’s 4k token limitation!

Live demo:

Covering some amazing applications, e.g., from a handwritten UI mock-up to functional code. Or complex taxation questions!

Bing AI Uses GPT4 tweaked for search

Hallucination and other limitations? what’s new?

Does it still hallucinate & is it factual?

Short answer, GPT4 is significantly more reliable than ChatGPT but not bullet proof. So per our usual advice: trust & verify!

Disclamer in OpenAI’s own words, it still hallucinates be careful: “Despite its capabilities, GPT-4 has similar limitations as earlier GPT models. Most importantly, it still is not fully reliable (it “hallucinates” facts and makes reasoning errors).”

YET, it is way more reliable: “GPT-4 scores 40% higher than our latest GPT-3.5 on our internal adversarial factuality evaluations”.

What about Safety and harmful content?

“GPT-4 poses similar risks as previous models, such as generating harmful advice, buggy code, or inaccurate information.”

Yet, GPT4 offers a better alignment with OpenAI’s goals to design a safer AI. It means limiting answers to sensitive topics or harmful behavior.

(source)

Confidence & calibration: Still some work to do!

This part hurts a bit!

When we rely on an AI model it’s important to know when the AI is confident in the answer and when it’s not. By being calibrated, we can also design thresholds to know when to reliably accept AI’s answer or not. Current post training of the model harms the calibration curve unfortunately.

So…GPT4 or ChatGPT?

Model evolution: GPT4 is clearly a huge leap forward compared to ChatGPT. Just look at these numbers.

With scores putting it in the top 10% of students across the board, it can pass the bar exam or integrate any prestigious university hands down!

GPT4 shows a huge improvement in reasoning tasks compared to ChatGPT
Yop scores for GPT4 on all exams!

Let’s answer the question based on an ROI evaluation.
GPT4 unlocks scenarios that where impossible before, yet…
Do you really really need it? Well the answer is “it depends”:

  • If image inputs can greatly enhance your result, then yes,
  • If you need a much much large context (i.e., more than 3000 words), then yes,
  • If having a model that is way more factual is crucial for your application, then definitely yes!

Otherwise,

  • If your application can live with a 4k tokens (3k words) context with simple reasoning, then ChatGPT has the best ROI choice.
  • If semantic search is enough to capture real time additional context, then ChatGPT should do the trick too!

Lean more

Want to know more about GPT4 inner mechanics? read this post about the 3 main optimizations that are supercharging it. Our predictions are aligned with GPT4’s alpha release, i.e., GPT4 is text only for now.

Want to start building with OpenAI APIs, these two posts get you covered.

Conclusion

In conclusion, the release of GPT-4 is a significant milestone in the development of AI language models. It is the first multi-modal model that can process both text and images as input, making it a powerful tool for various applications. GPT-4 is also much smarter than its predecessor, ChatGPT, particularly in complex reasoning tasks. However, access to the API is currently limited to a waitlist, and the prices are significantly higher than ChatGPT.

While GPT-4 is more reliable than ChatGPT, it still has limitations and can generate inaccurate information or harmful advice. Moreover, its calibration curve needs improvement, which affects its confidence and reliability. Therefore, it is essential to verify its output before using it in critical applications.

Whether to choose GPT-4 or ChatGPT depends on the specific application’s needs. GPT-4 is an excellent choice for scenarios that require image inputs, a much larger context, or factual accuracy. On the other hand, ChatGPT is still a good ROI choice for applications that can live with a 4k token context and simple reasoning or rely on semantic search.

Overall, GPT-4 is an impressive AI language model with significant potential for various applications. However, it is crucial to consider the cost, limitations, and specific application requirements before deciding to use it.

--

--

BoredGeekSociety

Wassim Jouini, CTO and Head of AI @LegalPlace. 12yrs+ Building AI & Automation products for Scale-Up Startups | Ph.D. AI