Perplexity AI

[backtothebeach]

In the last few days I have heard perplexity.ai being mentioned in various podcasts, and now here on the board in the Google thread.

 

The name sucks IMO, but this could catch on. Maybe they find a catchy name for their engine that could become a verb... Private company for now.

 

From Wikipedia:

 

"Perplexity was founded in August 2022 by Aravind Srinivas, Denis Yarats, Johnny Ho, and Andy Konwinski.^[7] Aravind Srinivas previously worked at OpenAI and now serves as Perplexity's CEO.^[8]
 

"Perplexity has raised $73.6 million in a Series B funding round, led by notable investors including Nvidia and Jeff Bezos. This investment, part of the total $100 million raised so far, has contributed to the company's current valuation of $520 million.^{[15][16][2][17][18]} The investors so far include New Enterprise Associates, Nvidia, Databricks, Bessemer Venture Partners, Jeff Bezos, Susan Wojcicki, Jeff Dean, Yann LeCun, Andrej Karpathy and others.^{[19][20][8][21][18]}

Despite this high valuation, Perplexity has fewer than 40 employees.^[19] It has garnered about 10 million monthly users.^[22]

[backtothebeach]

Q: who are the investors that participated in perplexity ai's first funding round

 

A: (by perplexity, removed footnotes)

Perplexity AI's initial funding round included investors such as New Enterprise Associates (NEA), Databricks Ventures, Elad Gil, Nat Friedman, Pieter Abbeel (AIX Ventures), Yann LeCun (Meta), Andrej Karpathy (OpenAI), Ashish Vaswani (Transformer), Amjad Masad (Replit), Clem Delangue (HuggingFace), and Oriol Vinyals (DeepMind). These investors were involved during the seed stage and the Series A funding round before the company gained significant recognition and value.
 

Edited February 25 by backtothebeach

[Luke]

I have used this extensively and it's excellent. You can choose between several language models, i am using GPT 4, its very precise, i can put data outputs in it and have it put out competent summaries, saving me tons of time. I can simply screenshot SPSS outputs as an example and it gives me very good write ups within a couple of seconds that are accurate most often...20 USD a month is a steal for the time saver it is.

 

And for other questions it gives me a better output than what google gives me...perplexity ai seriously replaced a significant part of my google searches.

Edited February 25 by Luca

[Luke]

It shocked me when i sent some screenshots of a rather difficult analysis SPSS output into the program and i received such a detailed correct write up...wtf man...these tools are so powerful.

Edited February 25 by Luca

[mattee2264]

Isn't this the worry really?

 

The basic technology behind AI is out there and not proprietary. Mag7 companies have advantages because they are stockpiling chips, have the biggest R&D budgets by a huge distance, have existing capabilities in related technologies such as data analytics, machine-learning etc. 

 

It may be though that the practical real-life applications in the near term at least might be more niche and the niche markets just do not move the needle for Big Tech and they are instead going to waste a lot of money on overly ambitious projects such as AI co-pilots which don't add enough value to companies to allow them to be able to recoup their massive investments in AI. 

 

And the whole appeal of companies like Microsoft, Meta, Google etc is they were so capital light and gushed FCF. Now all that FCF is diminishing because AI is very capital intensive with all the chips and so on. And they are cutting staff and risk neglecting their core businesses because they are seduced by the holy grail of AGI. 

 

 

[bargainman]

LeCun and Karpathy are luminaries in the AI field.  Yes LLMs are pretty impressive.  Multimodal models are too.  Lots of big challenges ahead but moving at breakneck speed.

[lnofeisone]

I'm also a fan of perlexity. Tried few things that are "hard" and got very impressive answers that saved at least 2-3 weeks of work. 

[Luke]

23 minutes ago, lnofeisone said:

I'm also a fan of perlexity. Tried few things that are "hard" and got very impressive answers that saved at least 2-3 weeks of work. 

yeah exactly! 

[Gmthebeau]

I have used it and was impressed also.  For a new product it is really good, will only get better.  

[ValueArb]

Perplexity seems like a perfect name for an AI tool.

 

Is it cool? Hell yes!

is it accurate? I’m perplexed!

[Hektor]

1 hour ago, ValueArb said:

Perplexity seems like a perfect name for an AI tool.

 

Is it cool? Hell yes!

is it accurate? I’m perplexed!

[nwoodman]

I have been using this for a while too.  It’s very difficult to go back to normal search, that’s for sure. The potted questions are a bit hit and miss at times but the referencing is brilliant.  For technical engineering searches it has proved invaluable and easily justifies its cost.  

For investment research it is OK to good.  It sometimes dives into filings but more often than not gives the usual lazy Chat GPT response eg.

 

“one would typically look into annual reports, SEC filings, and other detailed investor communications that outline the company's strategic investment decisions over time. However, these details are not provided in the current search results”

 

 

[ValueArb]

Here is the first query I gave Perplexity, asking for a list of public companies in the US that are net-nets. None of the tickers it gave me are trading, some went private 5 years ago, others I have no idea if they ever existed.

 

Quote

Give me a list of public companies in the US that are net-nets

 

Sources

 

 

 

View 2 more

 

Answer

Here is a list of public companies in the US that are net-nets:

• ELOS
• ACTS
• HLYS
• HRAY
• HTX
• IKAN
• LTON
• MEMS
• NED
• PDII

These companies are considered net-nets, which are undervalued stocks traded below their net current asset value (NCAV), calculated as Current Assets minus Total Liabilities

1

2

3

. Net-net stocks are often seen as deep value investments that can potentially offer significant returns

1

.

 

[formthirteen]

5 hours ago, ValueArb said:

Give me a list of public companies in the US that are net-nets

 

This is something these services are unable to calculate or answer at the moment. These large-language models are very good at interpreting human language and generating text. They are trained on huge amounts of text and have a ”faint memory” of almost everything they are trained on. They would need a ”large-number model” that has the ability to calculate this together with perfect recollection of the data, in other words, access to a database and a calculator. There are probably services like this being built already...

 

You could try asking ”Can you give me a list of net-net stock screeners?” or ”Can you give me a list of net-net stock screeners that are integrated with a large-language model?”.

[tnp20]

I hvave been using perplexity for the last 6 months. Its petty good. Specially its ability to provide source material links to verify its not halucinations.

 

Its also good because you can change the underlying model it uses for your prompt....I have the paid version so I can use GPT-4 turbo..there is auto mode where it picks the model for you.

 

Also if you can get access Anthropic and notebooklm.google.com are also good for PDF documents as input....

 

Use cases so far:-

 

(i) General search but more complex query 

 

(ii) Resume - asked it how good a match resume was to the job description and how to tweak it to get a better match

 

(iii) Financial chat board and message board summrization - some stocks have 100s of posts per day - I am able to print it to PDF and load it into one of the AI models and summarize it and also ask it more specific questions that may have been discussed

 

(iv) many others that I can't talk about

[ValueArb]

1 hour ago, tnp20 said:

I hvave been using perplexity for the last 6 months. Its petty good. Specially its ability to provide source material links to verify its not halucinations.

 

Providing sources is a very nice feature, and looking at mine its relying on web pages that are way outdated. I don't think its Perplexity's fault because the pages appear to be undated, So I tried modifying the verbiage to "that are currently net-nets" but get the same list back.

 

7 hours ago, formthirteen said:

 

This is something these services are unable to calculate or answer at the moment. These large-language models are very good at interpreting human language and generating text. They are trained on huge amounts of text and have a ”faint memory” of almost everything they are trained on. They would need a ”large-number model” that has the ability to calculate this together with perfect recollection of the data, in other words, access to a database and a calculator. There are probably services like this being built already...

 

You could try asking ”Can you give me a list of net-net stock screeners?” or ”Can you give me a list of net-net stock screeners that are integrated with a large-language model?”.

 

Thanks! will try this next.

Edited March 1 by ValueArb

[bargainman]

Well there's a whole new field of so called "Prompt Engineering", which talks about how to properly ask questions of the LLMs to get good answers.  if you are serious about using them you probably want to check that out.

 

This is a free short course for developers but might be useful for others

https://www.deeplearning.ai/short-courses/chatgpt-prompt-engineering-for-developers/

 

There are likely more suitable ones for the chat interfaces.

[ValueArb]

On 3/1/2024 at 5:59 PM, bargainman said:

Well there's a whole new field of so called "Prompt Engineering", which talks about how to properly ask questions of the LLMs to get good answers.  if you are serious about using them you probably want to check that out.

 

This is a free short course for developers but might be useful for others

https://www.deeplearning.ai/short-courses/chatgpt-prompt-engineering-for-developers/

 

There are likely more suitable ones for the chat interfaces.

 

We need an upvote mechanism on the site for posts like this.

[rogermunibond]

Isn't prompt engineering like Google-fu?

[nwoodman]

OT but Claude 3 just dropped and so was playing with it today.  I actually thought this wasn’t too bad in terms of output:

 

Sure, I'll provide a basic attempt at all 5 valuation methods for Fairfax Financial. Please note that these are simplified examples and should not be considered a comprehensive valuation or investment advice.

1. Sum-of-the-Parts (SOTP) Valuation:
   - Insurance and reinsurance subsidiaries: $20 billion (based on book value and peer comparisons)
   - Investment portfolio: $30 billion (based on market value)
   - Other assets and liabilities: -$5 billion
   - Total SOTP value: $45 billion

2. Comparative Valuation:
   - Fairfax's P/B ratio: 1.2 (current stock price / book value per share)
   - Peer average P/B ratio: 1.5
   - Implied valuation based on peer average: $52.5 billion (1.5 * book value)

3. Discounted Cash Flow (DCF) Analysis:
   - Estimated 5-year cash flows: $5 billion (based on historical growth and profitability)
   - Terminal value after 5 years: $60 billion (based on a perpetual growth rate of 2%)
   - Discount rate: 8% (based on the company's cost of capital)
   - Present value of cash flows and terminal value: $48 billion

4. Dividend Discount Model (DDM):
   - Current annual dividend per share: $10
   - Expected dividend growth rate: 5% (based on historical growth and payout ratio)
   - Cost of equity: 8% (based on the risk-free rate, market risk premium, and company-specific risk)
   - Implied value per share: $333 ($10 / (0.08 - 0.05))
   - Total implied valuation: $50 billion (based on the current number of shares outstanding)

5. Adjustments for Unique Factors:
   - Add a 10% premium for strong management and investment track record
   - Subtract a 5% discount for exposure to catastrophe losses and market volatility
   - Net adjustment: +5%

Based on these simplified examples, the valuation range for Fairfax Financial could be:
- Low: $45 billion (SOTP valuation)
- High: $55.1 billion (Comparative valuation with peer average + 5% net adjustment)
- Average: $50 billion

Again, this is a basic attempt and should not be considered a comprehensive valuation. A thorough valuation would require more detailed analysis and assumptions.

[backtothebeach]

Interview with Perplexity’s founder and CEO:

https://podcasts.apple.com/us/podcast/invest-like-the-best-with-patrick-oshaughnessy/id1154105909

[nwoodman]

9 hours ago, backtothebeach said:

Interview with Perplexity’s founder and CEO:

https://podcasts.apple.com/us/podcast/invest-like-the-best-with-patrick-oshaughnessy/id1154105909

Excellent podcast. A lot to unpack but listening to it only reinforced the old adage “I can’t afford free”.  I have money and are very happy to pay for a search/knowledge experience where it is serves results, that are high quality, more often than not delight and are easily verified.  Aravind Srinivas’, laser like focus, on UX, is indeed what makes this great.  The market can no doubt see the ramifications for Google and as he says:

 

“..the differentiation is going to come from us executing even faster and better, because unlike a big company, if you call them a big company (ref OpenAI), they're actually pretty fast. So we have to be even faster. So that is one thing I do think about. Unlike what most people say, I'm not really worried about Google, not because they cannot execute on this. They actually are way better engineers and researchers than us. It is their own business model.”

 

The other thought is that this is where Apple can shine and not just UX.  This is an energy intensive endeavour. “High quality answer”/watt in the medium term could  be a real differentiator. He touched on the importance of using relatively smaller, more efficient language models that are fine-tuned for specific tasks like avoiding hallucination, rather than depending on the largest, most complex models. He suggests that better "reasoning" capabilities in smaller models could be a key to making this approach more economical.  On device could make a big difference here IMHO. Not to mention the privacy considerations and personalisation opportunities..

 

What an amazing time to have a heartbeat.

Edited March 6 by nwoodman

[rohitc99]

On 3/1/2024 at 8:11 AM, tnp20 said:

I hvave been using perplexity for the last 6 months. Its petty good. Specially its ability to provide source material links to verify its not halucinations.

 

Its also good because you can change the underlying model it uses for your prompt....I have the paid version so I can use GPT-4 turbo..there is auto mode where it picks the model for you.

 

Also if you can get access Anthropic and notebooklm.google.com are also good for PDF documents as input....

 

Use cases so far:-

 

(i) General search but more complex query 

 

(ii) Resume - asked it how good a match resume was to the job description and how to tweak it to get a better match

 

(iii) Financial chat board and message board summrization - some stocks have 100s of posts per day - I am able to print it to PDF and load it into one of the AI models and summarize it and also ask it more specific questions that may have been discussed

 

(iv) many others that I can't talk about

have you looked at chatGpt or done comparison with perplexity ? any feedback ?

 

thanks in advance

[mattee2264]

Isn't it a little worrying for Mag7 investors that a start-up founded only a few years ago can release LLMs that compare favourably to the ones put out by OpenAI/Microsoft and Google? Reinforces the point that in AI no one really has a moat at this point. Mag7's ability to stockpile chips gives them an edge and they can outspend everyone on R&D. But in the internet age it was Google who ended up with the dominant search engine not AOL/Netscape/Microsoft. 

Any company with a good idea/product will have little problem attracting funding and users aren't locked in to a specific LLM at this point so will switch if something better comes along. 

 

And so far it is LLMs that are generating all the buzz and drawing in all the users and seeming to have the most practical use as a lot of people are using LLMs to write emails, assignments, research papers, marketing copy etc. 

[Spekulatius]

I don’t think the LLM itself have much value and any edge seems to be fleeting and probably only lasts a few month. It seems difficult to build a moat around an LLM alone. The moat if any will de slop around specific use cases, based on training on proprietary data and perhaps in developing an ecosystem that makes an LLM easier to use.