Chat GPT & Generative AI tools

[wearthefoxhat]

I read the paper on Deepseek and basically they cut a few corners to make it much cheaper, but quite effective. It's like taking a luxury car and removing lots of the bits but saying, well it still drives from A to B.

I think a lot of valuations are still frothy thought. So selloff overdue and maybe a moment for people to check they are still on the right path and exactly what they path will be.

Curiosuly Deepseek thinks it's ChatGPT if you ask it.

I wasn't far off then....

deepGPT.png

[ShaunWhite]

Who needs to spend $500 billion when you can access GPT for free....or am I missing something?...

It's definately not free. I'm paying $20 for 4o. o1 is $200/mo and have you seen the proposed price of o3? Low-compute is $20/task and high-compute mode is $1000/task.

[sionascaig]

Who needs to spend $500 billion when you can access GPT for free....or am I missing something?...

It's definately not free. I'm paying $20 for 4o. o1 is $200/mo and have you seen the proposed price of o3? Low-compute is $20/task and high-compute mode is $1000/task.

I saw an advert for a new Pixel phone which came with free AI support for the 1st year, then £40pm thereafter (that is £40pm just for the Gemini assistant)..

I really can't see any mass take up at those prices and perhaps some companies are deluding themselves..

You can get a flight to Morocco from Edin atm for £40!

[ShaunWhite]

The internet democratised knowledge but advanced AI will be firmly in the hands of those who can afford it.

[wearthefoxhat]

Who needs to spend $500 billion when you can access GPT for free....or am I missing something?...

It's definately not free. I'm paying $20 for 4o. o1 is $200/mo and have you seen the proposed price of o3? Low-compute is $20/task and high-compute mode is $1000/task.

There is a free version, although has limitations.

2 tiers.png

Saw something on twitter/X a while back about Abacus.

Abacus.png

https://chatllm.abacus.ai/gre

[Hepburn]

I tried to feed deepseek some cutdown data. Its said go away.

[sionascaig]

I tried to feed deepseek some cutdown data. Its said go away.

Yup.. Its under cyberattack atm... Maybe some folk have too much too lose...

At least we know its not the Chinese this time )

[ShaunWhite]

There is a free version, although has limitations.

I used the free one initially but the paid is so much better, esp at code debugging and analysing data. It also does a better job of understanding/suggesting strategy. I might treat myself to a month or two on o1 as I've heard good things about it.

[foxwood]

Plus gives you limited usage of o1 - toggle the dropdown at the top to change model - asked it what the limits were

what are limits in the Plus version of gpt chat to usage of the o1 model ?

o1 replied a very "abrupt" answer (settings are don't be verbose lol) ...

Thought about usage limits of GPT-4 for 16 seconds
Currently, ChatGPT Plus users can send around 50 GPT-4 messages every 3 hours. These limits may change over time.

4o was verbose and more sensible

The Plus version of ChatGPT, using the GPT-4 (GPT-4-turbo) model, generally provides enhanced performance and faster response times compared to the free tier. As of now, usage limits for the Plus plan include:

Message Cap:
There is typically a limit on the number of messages you can send per session or within a specific time frame (e.g., 50 messages every 3 hours). This may be subject to change.

Session Management:
Sessions might be rate-limited or throttled during periods of very high demand.

Token Limit:
GPT-4-turbo has a context window of approximately 128k tokens, enabling longer conversations or processing more extensive prompts and outputs. However, this is per session or conversation.

For precise and updated limits, refer to the official documentation or announcements from OpenAI, as these policies may evolve.

[Fugazi]

Who needs to spend $500 billion when you can access GPT for free....or am I missing something?...

It's definately not free. I'm paying $20 for 4o. o1 is $200/mo and have you seen the proposed price of o3? Low-compute is $20/task and high-compute mode is $1000/task.

Been excited about the future of AI purely from a BA / coding perspective.

No longer feels magical hah

[wearthefoxhat]

My first impression of DeepSeek for excel solutions was a good one.

I have an ongoing thread on GPT where I'm building a horse racing model from various pieces of data from other websites. I asked GPT about a line of numbers and letters that get brought across from the web as text, and the number is produces, is read as text by my excel file.

GPT in micro seconds produced: =IF(ISNUMBER(VALUE(RIGHT(AN5,1))), IF(RIGHT(AN5,1)="0", 10, VALUE(RIGHT(AN5,1))), 10)

DS in 2 seconds and produced: =IF(OR(RIGHT(AN5,1)="0", NOT(ISNUMBER(VALUE(RIGHT(AN5,1))))), 10, VALUE(RIGHT(AN5,1)))

Both worked well. (Free tiers) - Haven't got a spare $500 Billion that's for sure...(Can be done for a measly $6million by the sounds of it)

If I had to create that on my own...FFS, I'd be here until the next ice age trying to figure it out!

[ShaunWhite]

Both need to be wrapped in an =iferror(formula, default) expression.
GPT doesn't code defensively (unexpected data) unless you point out its not using best practice.

In fact when asked to critique those responses the paid version says.....

There's a critical flaw in both formulas: neither uses IFERROR to handle potential errors cleanly.

Critical Points:

1. Error Handling:

VALUE(RIGHT(AN5,1)) assumes RIGHT(AN5,1) will always be a valid number or "0".

If RIGHT(AN5,1) contains non-numeric text (e.g., a letter), VALUE will throw an error instead of defaulting to 10.

A better approach: IFERROR(VALUE(RIGHT(AN5,1)), 10), which ensures any error results in 10 rather than breaking the formula.

2. Efficiency:

The OR(RIGHT(AN5,1)="0", NOT(ISNUMBER(VALUE(RIGHT(AN5,1))))) check in DM’s version is unnecessarily convoluted.

A streamlined, robust formula would use IFERROR, simplifying both readability and execution.

Conclusion:

While both formulas work in common cases, they lack robustness. The over-reliance on ISNUMBER and VALUE without error trapping is a weakness. A well-structured Excel formula should anticipate potential errors instead of assuming clean inputs

[wearthefoxhat]

Both need to be wrapped in an =iferror(formula, default) expression.
GPT doesn't code defensively (unexpected data) unless you point out its not using best practice.

In fact when asked to critique those responses the paid version says.....

There's a critical flaw in both formulas: neither uses IFERROR to handle potential errors cleanly.

Critical Points:

1. Error Handling:

VALUE(RIGHT(AN5,1)) assumes RIGHT(AN5,1) will always be a valid number or "0".

If RIGHT(AN5,1) contains non-numeric text (e.g., a letter), VALUE will throw an error instead of defaulting to 10.

A better approach: IFERROR(VALUE(RIGHT(AN5,1)), 10), which ensures any error results in 10 rather than breaking the formula.

2. Efficiency:

The OR(RIGHT(AN5,1)="0", NOT(ISNUMBER(VALUE(RIGHT(AN5,1))))) check in DM’s version is unnecessarily convoluted.

A streamlined, robust formula would use IFERROR, simplifying both readability and execution.

Conclusion:

While both formulas work in common cases, they lack robustness. The over-reliance on ISNUMBER and VALUE without error trapping is a weakness. A well-structured Excel formula should anticipate potential errors instead of assuming clean inputs

Thanks.

[ShaunWhite]

It's OK, but would I trust it designing a nuclear reactor or a new medicine, nope.

[wearthefoxhat]

Add in DeepSeek and away you go.

AI 1b.png

AI 2.png