GPT 4

[formthirteen]

On 3/26/2023 at 3:20 PM, Spekulatius said:

How do you guys deal with false answers?

 

It's like having a smart and know-it-all assistant who you can't trust 100% (didn't take the meds or took the wrong ones). The assistant gets the job 80% right, and the rest you can hopefully complete yourself. Google has the same issue and you're stuck doing the job of the assistant who gathers information (from spam farms) and does the initial draft (better than you). ChatGPT gives me a productivity boost compared to Google. I'm happy to pay $20/month for ChatGPT an assistant.

 

How to improve the reliability of ChatGPT is explained in this official guide:

https://github.com/openai/openai-cookbook/blob/main/how_to_work_with_large_language_models.md

https://github.com/openai/openai-cookbook/blob/main/techniques_to_improve_reliability.md

 

Someone could also create an "air conditioner assistant" by fine-tuning a ChatGPT model for your use case, if they had the data (e.g. questions and answers):

https://platform.openai.com/docs/guides/fine-tuning

https://github.com/openai/openai-cookbook/blob/main/examples/Fine-tuned_classification.ipynb

Edited April 4, 2023 by formthirteen

[Spekulatius]

LOL, we are all going to be rich.

This chatGPT is so bad in math, it’s really hard to believe. I also have seen it make errors all over the lace

 

Apparently according to ChatGPT

100/1= 1000

 

and

 

1000^0.1= 1.07024

 

Its going to be a lot of fun when students use this for homework.

 

 

 

[RedLion]

Started playing with this today and think it’s going to be very useful in my work to delegate low level tasks. 

[competitive-advantage]

@tnp20 Thanks for your experiences. It's really helpful 

 

The astonishing realization I had is that most questions can be answered by asking ChatGPT. While I don't always agree with its responses, it often provides a solid starting point for further research. I concur with you, @Spekulatius , that ChatGPT can occasionally be accurate, as demonstrated by the teenage example. It's crucial, however, that we offer feedback through upvotes/downvotes and written comments, and even engage directly with ChatGPT.

 

It was an intriguing experience creating the comedic post above with ChatGPT. Although the AI generated the text, I provided substantial feedback and brainstorming. However, when I mentioned our collaboration in creating comedy, ChatGPT kept insisting that the AI did the heavy lifting and my contribution was minimal. I later realized that, of course, the AI needs feedback as it doesn't yet fully grasp the extent of human collaboration required.

 

I have been using ChatGPT extensively for a food website I am developing. The AI excels in areas like writing, strategizing, coding, debugging, and speed. It feels like working in a team of 3-4 people instead of just by myself. When in doubt, I can ask ChatGPT and usually receive helpful guidance, similar to what I might get in a department meeting.

 

As for accessing the latest data, there's a Chrome plugin that allows internet access, but it isn't perfect. Alternatively, you can provide ChatGPT with text from earnings calls, for example, and then ask questions. It will be fascinating when ChatGPT can actually remember things, as currently, I have to repeat information when a new day begins.

 

I also see potential for AI in mental health support. If someone is struggling and lacks a friend to talk to, they can engage in conversation with AI. While it's not human, it could still provide valuable assistance to many people.

Furthermore, ChatGPT is a game-changer for individuals who are not fluent in English (or other languages) or have dyslexia.

 

 

[formthirteen]

19 hours ago, Spekulatius said:

This chatGPT is so bad in math, it’s really hard to believe. I also have seen it make errors all over the lace

 

https://writings.stephenwolfram.com/2023/03/chatgpt-gets-its-wolfram-superpowers/

[frommi]

Whats the moat here? In 5 years a lot of companies will have these type of language models, but it requires a lot of capital to run because computing hours are expensive. Maybe shorting some companies is successful as an approach, but its a low return endeavour and very dangerous. I doubt that GOOG or MSFT will ever really make money of it. But of course when first playing around with it i was fascinated, too. 

[competitive-advantage]

4 hours ago, frommi said:

Whats the moat here?

 

The competitive edge emerges when the superior AI for task X draws in a majority of clients. As a company acquires more customers, its performance enhances, leaving competitors struggling to keep up.

 

I employ AI for creating content and discovered Chat GPT to surpass Jasper AI in terms of capability. Furthermore, Chat GPT excels in formulating strategies, programming, and addressing diverse queries.

 

For businesses, whether the AI replaces 1, 2, or 50 writers, the monthly expense of $1,000 or $5,000 is inconsequential.

[Spekulatius]

15 hours ago, competitive-advantage said:

 

The competitive edge emerges when the superior AI for task X draws in a majority of clients. As a company acquires more customers, its performance enhances, leaving competitors struggling to keep up.

 

I employ AI for creating content and discovered Chat GPT to surpass Jasper AI in terms of capability. Furthermore, Chat GPT excels in formulating strategies, programming, and addressing diverse queries.

 

For businesses, whether the AI replaces 1, 2, or 50 writers, the monthly expense of $1,000 or $5,000 is inconsequential.

I think AI bots will be a supplementary business for cloud providers basically. That means MSFT, AMZN, GOOGL maybe Alibaba in China. They already have the server and computing capacity.

[patience_and_focus]

1 hour ago, Spekulatius said:

I think AI bots will be a supplementary business for cloud providers basically. That means MSFT, AMZN, GOOGL maybe Alibaba in China. They already have the server and computing capacity.

This.

 

IMHO somewhat similar to iPhone /Android apps ecosystem that decimated a lot of niche businesses, this will replace a lot of them. But the lions share might go towards the top dogs like Google, MSFT, etc.

 

Apple is so far a non player it seems like, unless there is a secret internal effort or they buy the talent and product by acquisition.

[tnp20]

>> Whats the moat here? In 5 years a lot of companies will have these type of language models, but it requires a lot of capital to run because computing hours are expensive. Maybe shorting some companies is successful as an approach, but its a low return endeavour and very dangerous. I doubt that GOOG or MSFT will ever really make money of it. But of course when first playing around with it i was fascinated, too.  <<

 

Good questions. From my experience, extrapolated....

 

(i) My wife is a senior executive for a fairly big consulting company. A lot of her big clients are already thinking about it and starting to act...we are talking BIG $$$$ commitments here....her current client is a fortune 100 global company, its not an I.T. company, they are thinking of transforming their I.T. to take advantage of what AI can offer both today and tomorrow....

 

(ii)  After trying out chatGPT for one night, I signed up for $20/month chatGPT 4 model...there are many like me who believe the value to AI is immense ...perhaps bigger than the internet...Bill Gates has said this is the next big revolution like the internet and when you play with it you realize why its true. Its not just today, but what it will be in 5-10-20 years. Today's its capabilities are incredibly impressive...scarily so in many cases....its a huge productivity tool...

 

(iii) If everyone deploys AI, where is the edge ?

                  - There will be an edge to those who have the best a AI systems, models, complexity, balance of trade offs and cut offs

                  - Uniqueness of data that is used to train ...if you have special data that no one else has ...your models are going to produce some incredibly valuable intelligence

                  - How well the models are trained with the data in terms of diversity of data, the method of training etc.

                  - How AI is deployed optimally to utilize cost/benefit ratio most effectively

 

(iv) Cloud usage will explode because of AI ....AWS, Azure, Google cloud, Tencent, BABA, Baidu will do well over the next 20 years.

 

If you are not both excited and scared by most recent incarnation of AI( chatgpt, bard, Watson, Lexa, Boston gynamics, Waymo etc....) you have not understood AI....

 

Edited April 9, 2023 by tnp20

[Intelligent_Investor]

So my company's version of MS Teams now has an official GPT-4 Virtual Assistant. Pretty powerful in terms of what you can do with it

[Saluki]

My better half works at McKinsey, and she is way more tech literate than I am.  She mentioned that she just finished a course on ChatGPT on Udemy which was helpful.  It helps to learn how to give better prompts and use prior answers to shape future queries apparently.  For instance, she had it generate lists of best cities to live with these 5 criteria and rank them in an excel sheet (I didn't know it could do things besides write bad poetry and give passable, generic history essays for high school students). 

[tnp20]

Another way to think about AI market in the future is (i) Custom/Niche and (ii) Existing products

 

If we look at existing products and services like cloud infrastructure, search, Office Suites, ERP, CRM, EMR

Easiest way to push out AI to customers is through existing products by incorporating enhanced functionality and up charging.

 

1) Cloud: Amazon, Microsoft, Google, IBM ...in that order

2) Search: Google, Microsoft

3) Office Suites: Microsoft, Google, Adobe

4) ERP: SAP, Oracle, Microsoft (Ranking changes -so not sure on the order). 

5) CRM: Salesforce, Oracle, Microsoft

6) EMR: Oracle(Cerner acquisition), Epic, 

7) Mobile Users/Home devices: Apple, Google, Amazon

8 ) Hardware: Nvidia, Intel, Google, IBM, AMD

 

If we look at Niche markets where companies develop their own AI solutions because one is not available off the shelf, my best guess is Google, Microsoft, Amazon, IBM ...though the ranking will change over time based on execution and specific offerings based on their market advantages.

 

I think before companies themselves start using AI internally en-mass for niche/custom needs, its more likely they will buy AI augmented products from vendors above. Its hard to swap out ERP, CRM, EMR system - the costs are huge and adding AI entrenches the product even more in a company allowing for stickiness and pricing edge.

 

From investing point of view, its a lot of factors. Whats the size of the company and its share price and which has the most potential 10 years out in terms of share price growth. 

 

One mention...avoid C3.AI ..its changed its name 3 times....C3 Energy --> C3 IOT ---> C3.AI ...what ever flavour de jour ...

[tnp20]

I have been playing with ChatGPT, SonicChat and Bard. There is also Jasper but assholes want my credit card for a free trial and I am not up for that. (Jasper is well within its right but it likely wont get many early users and network effect.) 

 

Note this is personal experience and your mileage may vary.......I am just making personal observations like it or loath it !!!!!

 

Worst: Bing Chat...just not very powerful ...just does basic search and spits out what it can find on the web and its meaning......

 

Best:

 

Well it depends ....in order of current preference...

 

BARD: Logic, Economics, Finance, Mathematics (its got Minerva AI built in) = > , its both up to date and more accurate ....my go to place. Its not as verbose for professional writing like chatGPT.

 

SonicChat:   Pretty good because it has different modes such as:

 

 

and different libraries to do different things:-

 

 

It uses GPT4 underneath and it sites source material and you can tell it (prompt it) for "Re-write entire content that passes A.I. Tools Detection Test and give you highly-optimized, human written content." etc ...it seems it has lots of fine tuning mechanism over chatGPT even if it uses GPT4+ plus other AI models underneath.

 

ChatGPT4+:   Good for emails, presentations and asking about professional/consulting/business topics to write

 

 

 

 

 

 

Edited April 24, 2023 by tnp20

[ValueArb]

On 4/1/2023 at 10:00 AM, competitive-advantage said:

@tnp20 Thanks for your experiences. It's really helpful 

 

The astonishing realization I had is that most questions can be answered by asking ChatGPT. While I don't always agree with its responses, it often provides a solid starting point for further research. I concur with you, @Spekulatius , that ChatGPT can occasionally be accurate, as demonstrated by the teenage example. It's crucial, however, that we offer feedback through upvotes/downvotes and written comments, and even engage directly with ChatGPT.

 

It was an intriguing experience creating the comedic post above with ChatGPT. Although the AI generated the text, I provided substantial feedback and brainstorming. However, when I mentioned our collaboration in creating comedy, ChatGPT kept insisting that the AI did the heavy lifting and my contribution was minimal. I later realized that, of course, the AI needs feedback as it doesn't yet fully grasp the extent of human collaboration required.

 

I have been using ChatGPT extensively for a food website I am developing. The AI excels in areas like writing, strategizing, coding, debugging, and speed. It feels like working in a team of 3-4 people instead of just by myself. When in doubt, I can ask ChatGPT and usually receive helpful guidance, similar to what I might get in a department meeting.

 

As for accessing the latest data, there's a Chrome plugin that allows internet access, but it isn't perfect. Alternatively, you can provide ChatGPT with text from earnings calls, for example, and then ask questions. It will be fascinating when ChatGPT can actually remember things, as currently, I have to repeat information when a new day begins.

 

I also see potential for AI in mental health support. If someone is struggling and lacks a friend to talk to, they can engage in conversation with AI. While it's not human, it could still provide valuable assistance to many people.

Furthermore, ChatGPT is a game-changer for individuals who are not fluent in English (or other languages) or have dyslexia.

 

 

 

 

 

 

I work for decent sized Tech Unicorn and ChatGPT has been banned here. The issue is that to get the most out of it for development you need to feed it your existing codebase and ours is highly proprietary and no one wants to allow our competitors to catch up just because they were able to as ChatGPT for solutions our code solves. They are working on a solution right now, but no idea how long until we can get comfortable with it.

 

I'm also convinced it will have a valuable place in mental health. Psychology is a soft science whose studies have a lot of reproducibility issues. So the best way to provide therapy is unclear (and regularly changing). But there is reasonable support for the idea that just being able to talk to someone privately about your problems and issues improves perspective and mental outlook, so why not provide a bot at a fraction of the cost of a real therapist?

[tnp20]

SonicChat seems to have understood the mental health / psychology part of it....just look at the different modes...(i) Relationship coach (ii) Motivational coach (iii) Career counselor (iv) Personal Trainer (v) Stand-up comedian and the pièce de ré·sis·tance (vi) Philosopher mode for big picture life.

 

[competitive-advantage]

2 hours ago, ValueArb said:

 

 

I work for decent sized Tech Unicorn and ChatGPT has been banned here. The issue is that to get the most out of it for development you need to feed it your existing codebase and ours is highly proprietary and no one wants to allow our competitors to catch up just because they were able to as ChatGPT for solutions our code solves. They are working on a solution right now, but no idea how long until we can get comfortable with it.

 

Good idea to be cautious. I've talked with Chat GPT about what other users were using AI for and I got some pretty honest answers. Not specific details, but at least the subjects and it could potentially harm a companies competitiveness.

 

2 hours ago, ValueArb said:

 

I'm also convinced it will have a valuable place in mental health. Psychology is a soft science whose studies have a lot of reproducibility issues. So the best way to provide therapy is unclear (and regularly changing). But there is reasonable support for the idea that just being able to talk to someone privately about your problems and issues improves perspective and mental outlook, so why not provide a bot at a fraction of the cost of a real therapist?

 

That will be interesting. Especially if it really helps or it's just like social media where you get some good things, but many bad things as well. I can imagine it would be good if you ask for advice, get good advice and act on it. Otherwise I guess it's not good to rely on AI as a friend instead of training social skills.

[competitive-advantage]

1 hour ago, tnp20 said:

SonicChat seems to have understood the mental health / psychology part of it....just look at the different modes...(i) Relationship coach (ii) Motivational coach (iii) Career counselor (iv) Personal Trainer (v) Stand-up comedian and the pièce de ré·sis·tance (vi) Philosopher mode for big picture life.

 

 

It's interesting that you mention this. Because one thing is what the AI is capable of. Another thing is how user-friendly it is. It would be possible to imagine one AI being the best at doing hard tasks and another one being more user-friendly. The first AI could get some technical wizards excited, but the user-friendly AI will win the masses. It's happened before 

[tnp20]

https://finance.yahoo.com/news/amazon-cloud-business-just-had-013254201.html

 

 

[cameronfen]

Quote

I work for decent sized Tech Unicorn and ChatGPT has been banned here. The issue is that to get the most out of it for development you need to feed it your existing codebase and ours is highly proprietary and no one wants to allow our competitors to catch up just because they were able to as ChatGPT for solutions our code solves. They are working on a solution right now, but no idea how long until we can get comfortable with it.

 

I mean the easy solution is an on premises version.  Someone has to be working on that.  Seems like low-hanging fruit.  

[Morgan]

I don't have much to add, but it is clear that LLMs are going to change many things going forward. Check out the nearly hourly new tools and threads on Twitter. Just a few obvious examples: 

 

- Therapists will be replaced by LLMs at much lower price points. 

- Lawyers, doctors, assistants, accountants, etc will all need to adopt this into their organizations/workflow or they'll simply be less competitive. 

- Tons more junk will be published on the web as the cost of writing will go down by 90% (99%?).

- AI voice and video scams will become hugely more prevalent and costly to society. Example 1. Example 2. 

 

ChatGPT 4 was released 6 weeks ago, and already there are tons of new tools for a wide range of things. AI is evolving so fast! 

 

[tnp20]

[Spekulatius]

@tnp20 Do , if I believe that the magic quadrant assessment is correct, than IBM Is seriously undervalued with a market cap is just $115B vs trillions for competitors in the same quadrant.

[cameronfen]

52 minutes ago, Spekulatius said:

@tnp20 Do , if I believe that the magic quadrant assessment is correct, than IBM Is seriously undervalued with a market cap is just $115B vs trillions for competitors in the same quadrant.

Yea, this magic quadrant is ridiculous.  How do you put Clarifai, and H20.ai in the same discussion as Google.  There are 1000s of AI and machine learning firms that have more market power and broader vision than those two firms.  

[ValueArb]

On 4/28/2023 at 12:27 PM, cameronfen said:

 

I mean the easy solution is an on premises version.  Someone has to be working on that.  Seems like low-hanging fruit.  

100% agreed. But the effective models are huge and require mass scale in service and storage. I am a developer but not an AI expert so I don’t really understand how to solve the problem.
 

Can you build a proprietary model with your own data, that relies on the huge public model to fill in gaps in your proprietary model?  Or can it extend the public model without writing back to it?