Have We Hit The Top?

[lnofeisone]

39 minutes ago, ValueArb said:

Getting back to my legal example, feed your model all of the laws of your state, case histories, common law, precedents, etc, but don't let anything else contaminate it from the world wide web, then work with a team of paralegals and lawyers to come up with a strong testing regime to find and fix any mistakes, and maybe you have something that works as well as asking any lawyer who passed the bar. That would be hugely valuable product for the average joe in your state.

Great idea but it exists already - it's called Retrieval-Augmented Generation. There is also a heavy shift from general closed models (think chatGPT) to more open (so web + research articles) to more domain specific (e.g., specific area of law, specific area of tax code, etc.).

 

You also don't need a team of paralegals and lawyers if you are trying to help the "average Joe." Just need court data (verdicts, etc.) which is generally easy to get. 

 

Where you need a team of lawyers and actual law firms is some of those private settlements. I know of one law firm that has experimented with that but with mixed results. I have theories on why they had mixed results and happy to share them for a substantial fee (in case someone on this board is from that firm). 

 

Broadly speaking, however, AI is driving a lot of investments at the corporate level. Many of my clients are finding money for pilots, prototypes, and near-production-ready solution testing. Where a lot of them hit the pause button is when they realize that they need to revamp the entirety of their data ecosystems and project estimates go 10-100x but the willingness to experiment hasn't abated. 

[mattee2264]

If models need a lot of customisation or in-house programming to be useful wouldn't that limit the revenue opportunity for Big Tech? 

 

Presumably the idea is to provide for a subscription fee either the infrastructure or an off-the-shelf model that can easily be customized or programmed. And for the most basic applications/models there will probably be open source versions that allow companies to bypass Big Tech and allow niche smaller providers to make models for limited uses. 

 

OpenAI is a bit of a sham as they've already got into bed with Microsoft and clearly have little interest in making the underlying technology freely available for the benefit of mankind. But other Big Tech companies have come up with their own models and presumably lots of start ups can as well. 

 

Competition would drive down prices and returns and ultimately benefit consumers the most. 

 

But the way Big Tech companies have melted up since the release of Chat GPT there seems to be the implicit assumption that Big Tech will enjoy a lot of additional revenue with the same high margins and returns on investment they have made on their existing products/services. 

 

And maybe the internet is an example. It wasn't the companies providing access to the internet who made hay. But rather the companies who found a way to make money through the internet and even then there was a time delay and they weren't spared from the shakeout and even very big companies who were frontrunners didn't end up being the big winners. Only Microsoft is still on top. 

 

This time round Big Tech have a competitive edge as they have the huge customer base, the customers trust their products/services, they can invest far more than anyone else in AI and fund promising start ups and VC firms and reap the benefits of their work (and exclude others from doing so) but if the products they provide only provide a limited value add they won't be able to charge exorbitant fees and it may be that other smaller companies come up with more innovative uses and applications. 

 

 

 

[ValueArb]

19 hours ago, lnofeisone said:

You also don't need a team of paralegals and lawyers if you are trying to help the "average Joe." Just need court data (verdicts, etc.) which is generally easy to get. 

Without quality control, your outputs will be untrustworthy and useless.

[lnofeisone]

4 hours ago, ValueArb said:

Without quality control, your outputs will be untrustworthy and useless.

Judges and clerks who wrote decisions are your quality control. Lots of work has already been done by them, and incorporating them into your learning loop is essentially free. 

[cubsfan]

^^ A great story, and testament to CoBF. Well done Prasad & Dealraker.

[Gregmal]

Meanwhile Jerome Fauci just can’t let go of the spotlight.

[ValueArb]

12 hours ago, lnofeisone said:

Judges and clerks who wrote decisions are your quality control. Lots of work has already been done by them, and incorporating them into your learning loop is essentially free. 

 

That's not how pattern matching models like ChatGTP work to my knowledge (but please correct if you think I'm wrong, I'm an AI amateur who has only built models using images for minor projects, not in real work). But my long experience in software development leaves me skeptical about how easy it will be to turn AI models into trustworthy products.

 

The model will likely sometimes return the wrong decisions with the legal question given it. In the case of ChatGTP it will hallucinate imaginary decisions that never existed, but even if we constrain our model input set to proper legal databases instead of the World Wide Whacky Web that doesn't mean the pattern matching will always pick the correct decisions (or that it won't hallucinate non-existent ones). We've already seen ChatGPT written briefs get caught by judges, mostly because of hallucinations. 

 

This is the problem with all software, not just AI models. Before you implement any program, you have to have a strong quality control and testing plan for how to verify that it works well enough for production release. Obviously testing is almost always imperfect, how much you have to do and how well you do it is driven by product requirements. In the case of an "AI" legal advise bot, the potential costs of bad output are high (not as high as a medical bot, but still high) and you need to have a way to verify that questions are answered correctly almost all the time. That's why I suggest that building one should be a on a state by state basis (to avoid contamination and hallucination from other states different law sets), and output should be rigorously tested with actual legal experts.

[lnofeisone]

12 minutes ago, ValueArb said:

 

That's not how pattern matching models like ChatGTP work to my knowledge (but please correct if you think I'm wrong, I'm an AI amateur who has only built models using images for minor projects, not in real work). But my long experience in software development leaves me skeptical about how easy it will be to turn AI models into trustworthy products.

You have to deconstruct the problem. Usually LLMs are trained with question-answer pairings (take a look at SQuAD). It's what data you use and how you decompose it into Q-A pairs. Court cases tend to have a lot of reasons AND (this the core value of court data) you have a quantitative measurement of truth. For example, Party A sues Party B and asks for $1,000. If Party A gets $1000, you can label their argument as correct and Party B as incorrect (I'm grossly oversimplifying here). Party A gets $500, you can label it as half correct, etc. You can refine your vector database with assigned probabilities and use that to do dynamic Q-A build-outs. This works really well for general public use cases such as landlord disputes, small claims, etc. Doesn't work as well for big one-of-a-kind cases that will have major discontinuity (e.g., Roe v. Wade, Plessy v Ferguson -- Brown vs. Board of Ed). I would also hire a real lawyer if I have serious $$$ on the line. 

 

ChatGPT is trained with a lot of data, some of which can be questioned all day long. Court data is FAR better. It gets even better once you take into account metadata (e.g., timelines, law firms, etc.). Legal language is unique, though decisions have a bit of narrative, and can be somewhat arbitrary. General language models like ChatGPT or Gemini will not do well (due to hallucination for example) but if you decompose it into sub-models your results vastly improve. You can even get substantial result improvements with basic fine-tuning. If you can, play around with BloombergGPT. Few Gov't agencies are also building out their own LLMs, and the results are substantially better than ChatGPT/BARD, which is expected as those models are specialized (i.e., you'll get HORRIBLE answers if you ask something general but amazing results that are specific to the domain of the model). 

[changegonnacome]

1 hour ago, Gregmal said:

Meanwhile Jerome Fauci just can’t let go of the spotlight.

 

 Jerome Fauci - love it

 

But in all seriousness - he does have a problem - people are extrapolating the progress on inflation to date out to the 2% target......the reality is that we are hitting this stubborn underbelly of 3.x% inflation that isn't budging from what I can see (absent economic weakness which just isnt there)....productivity gains to the upside have pushed this persistently high inflation from my previous estimate of high 3's 'sticky' inflation to more like low 3's.

 

Not too sure Jerome has much room to cut - maybe one as a signal of strength/progress.....on the flipside productivity has surprised to the upside (I got this totally wrong).

 

its very interesting - this productivity progress - likely goes to show that American business is/was running with a tonne of fat on the bones - there was most definitely a churn in human capital in 2022/23.......and the re-allocation of talent from say Twitter's canteen/meditation rooms forcing these folks to get 'real' jobs definitely led to a productivity boost.

 

Ironically the 'crisis' at the Southern border - is no crisis at all......its a symptom of the immigration tap being turned on again (post-COVID) which feeds low end unskilled workers into an economy that desperately needs it (cooling inflation). The crisis at my southern border is not being able to hire someone to cut my lawn....or having to wait for months to get my deck re-done cause the decking company cant find workers anywhere!

[gfp]

7 minutes ago, changegonnacome said:

But in all seriousness - he does have a problem - people are extrapolating the progress on inflation to date out to the 2% target......the reality is that we are hitting this stubborn underbelly of 3.x% inflation that isn't budging from what I can see

 

Bro, it's mission accomplished

 

[Gregmal]

As I’ve said before I really don’t care if it is March or June or 3 or 30 rate cuts. It doesn’t impact how I invest. What’s perplexing is like Fauci, how this scumbag is basking in the spotlight, doing 60 minutes interview, etc. All his bros are placing wagers on recessions and he’s still talking about inflation lol. Both dumb and attention seeking. Same as the expert Fauci who totally made up 6 ft social distancing, told us vaccines stopped the spread, and enjoyed dousing himself in makeup prior to the television appearances. 

[Gregmal]

11 minutes ago, gfp said:

 

Bro, it's mission accomplished

 

Yea but Jerry heard some bartender in Kentucky got a 12% raise so now he’s worried about a wage price spiral. 

[ValueArb]

1 hour ago, lnofeisone said:

You have to deconstruct the problem. Usually LLMs are trained with question-answer pairings (take a look at SQuAD). It's what data you use and how you decompose it into Q-A pairs. Court cases tend to have a lot of reasons AND (this the core value of court data) you have a quantitative measurement of truth. For example, Party A sues Party B and asks for $1,000. If Party A gets $1000, you can label their argument as correct and Party B as incorrect (I'm grossly oversimplifying here). Party A gets $500, you can label it as half correct, etc. You can refine your vector database with assigned probabilities and use that to do dynamic Q-A build-outs. This works really well for general public use cases such as landlord disputes, small claims, etc. Doesn't work as well for big one-of-a-kind cases that will have major discontinuity (e.g., Roe v. Wade, Plessy v Ferguson -- Brown vs. Board of Ed). I would also hire a real lawyer if I have serious $$$ on the line. 

 

ChatGPT is trained with a lot of data, some of which can be questioned all day long. Court data is FAR better. It gets even better once you take into account metadata (e.g., timelines, law firms, etc.). Legal language is unique, though decisions have a bit of narrative, and can be somewhat arbitrary. General language models like ChatGPT or Gemini will not do well (due to hallucination for example) but if you decompose it into sub-models your results vastly improve. You can even get substantial result improvements with basic fine-tuning. If you can, play around with BloombergGPT. Few Gov't agencies are also building out their own LLMs, and the results are substantially better than ChatGPT/BARD, which is expected as those models are specialized (i.e., you'll get HORRIBLE answers if you ask something general but amazing results that are specific to the domain of the model). 

This was great, thanks!

[gfp]

41 minutes ago, gfp said:

 

Bro, it's mission accomplished

 

 

My biggest forward looking concern about inflation is a Trump presidency.  Trade wars, tariffs, deglobalization, tighter immigration.  I would add replacing Jay Powell if I thought the Federal Reserve had anything to do with inflation.

[Dinar]

15 minutes ago, gfp said:

 

My biggest forward looking concern about inflation is a Trump presidency.  Trade wars, tariffs, deglobalization, tighter immigration.  I would add replacing Jay Powell if I thought the Federal Reserve had anything to do with inflation.

Actually it depends on what Trump does.  If he does everything that you say, but guts social programs and eliminates income tax for singles with income under say $50K, $100K if single+kids, and say $100K for married couple, and $200K for married couple with kids, then you will have massive increase in the labor force keeping a lid on wages, and much lower budget deficit as social spending plummets while GDP grows sharply, and federal income taxes increase.  

Keep in mind, immigration is a massive drag on the economy right now.  Look how much NYC is spending on illegal immigrants - billions a year.  

Edited February 5 by Dinar

[Cod Liver Oil]

10 minutes ago, Dinar said:

Keep in mind, immigration is a massive drag on the economy right now.  Look how much NYC is spending on illegal immigrants - billions a year.

Mike, This is a big number!!!! Can you post a link to the cost? I am confused about our immigration policies.

Do we have any? What is the rationale?

[gfp]

Immigration might be a drag on certain state and local economies but it is a positive for the national economy.  Especially when you have labor shortages, low birth rates, undesirable demographic issues.  

 

[cubsfan]

33 minutes ago, gfp said:

 

My biggest forward looking concern about inflation is a Trump presidency.  Trade wars, tariffs, deglobalization, tighter immigration.  I would add replacing Jay Powell if I thought the Federal Reserve had anything to do with inflation.

 

Closing the Southern Border will be the key issue for the election. The current administration is killing the black community - and they are furious that their social services, parks, hospitals, schools, etc are being destroyed by President Biden. When Biden loses the black community and the southern states that are getting killed by immigration - they are going to look for a change to this disaster.

Edited February 5 by cubsfan

[Red Lion]

33 minutes ago, Dinar said:

Actually it depends on what Trump does.  If he does everything that you say, but guts social programs and eliminates income tax for singles with income under say $50K, $100K if single+kids, and say $100K for married couple, and $200K for married couple with kids, then you will have massive increase in the labor force keeping a lid on wages, and much lower budget deficit as social spending plummets while GDP grows sharply, and federal income taxes increase.  

Keep in mind, immigration is a massive drag on the economy right now.  Look how much NYC is spending on illegal immigrants - billions a year.  

I’d support Dinar for treasury secretary  

[thepupil]

5 minutes ago, RedLion said:

I’d support Dinar for treasury secretary  

From henceforth, 80-90% of you all won’t pay income taxes…lol

[changegonnacome]

1 hour ago, gfp said:

 

Bro, it's mission accomplished

 

 

Bro - its a website called Truflation....have you looked into it or do you just like the confirmation bias its feeding you...have a deeper look at truflation.com......its a bunch of blockchain web 3.0 hocus pocus bullshit attached...it has a whitepaper...they are literally selling shitcoins TRUF.....connected to the service.

 

Like look at this crapola:

 

Please dont send me crypto websites with proprietary undisclosed inflation models ....which at the same time are trying to sell unregistered securities linked to the service while simintoulsy promoting various crypto schemes. It is not a robust source of information.

 

 

 

 

[gfp]

So... not a fan of truflation  got it!

[changegonnacome]

3 minutes ago, gfp said:

So... not a fan of truflation  got it!

 

Not a fan of websites - selling token blockchain nonsense - while masquerading as some final authority on marco economic data.....you'd get thrown out of an economics 101 junior high school class for citing truflation in an essay....so not really interested in what the people at truflation have to say about anything really.

[gfp]

Well you wouldn't get thrown out of my class 

[Gregmal]

I’d take getting thrown out as a badge of honor given those courses and their ideology ultimately breed the disgusting legions of academically oriented politicians and bankers thatve fucked so much of society. 
 

It seems pretty clear that everything has normalized. I don’t need some academic chart to tell me what I see in day to day life. Even the energy guys have finally shut up.