Artificial Intelligence

[elliott]

1 hour ago, Spekulatius said:

Yes, apps already have API. Thats how they interact with other applications, at least the modern ones. I think any of the LLM agents will be build by the SaaS companies who build the current apps l it’s their market to lose.

 

Some of these AI maximists think they can just extract all the data over API’s and put the business logic in agents and the SaaS companies would just allow them to do that. First of all, business logic isn’t that simple and second the SaaS companies won’t give up their turf without a fight and taking a first shot at creating and selling the agents themselves.

 

Correct, SaaS companies are building agents and they will certainly fight.

But I can tell you that data from the applications is already being extracted...

Where I work, for example, is a large international company, and due to inorganic growth, there is a proliferation of applications, for example, multiple SAP systems. Data from all those SAP systems is being extracted every single day (MARA, MARC... all those tables), then integrated into a single respository, which many teams downstream are using for their own purposes. Agents that read this data are also being built and connected to the company's LLM. I don't know how it works, really, but with our LLM (just as when you use gemini or chatgpt) we can already ask questions related to our business data. E.g. what is the status of the item 00010 of order 1000056548?

 

Again, this is not the same as saying that SAP, or other SaaS, will be phased out easiliy. To begin with, the company where I work is not doing all this to phase out SAP, by any means. But you can see that things are evolving, and the future may not look like the present.

Edited March 1 by elliott

[frommi]

For reading/querying the data that is fine, but writing to it is a completly different story.

[nsx5200]

 

So it looks like the hackers are already on it.  Data leak through AI LLMs even with all the guard rails in place.

 

Like all the previous new 'features' before AI, everybody rush to adopt it, and then think about the security implications later (ex.  weak WEP encryption in previous the first generations of WIFI).  Not saying it can't be solved, but it does seem like this require a new set of techniques and time to shake out these 2nd order issues.

 

In other news, I heard in the conspiracy grape vines that China is seeding content on the open web with AI slop that's favorable for the CCP in order to get future LLMs to train on this and become more CCP friendly.  The long-term implication of this is that verified contents generated, or at least hand-checked by human will be valued more than output from AI systems, and rewarded accordingly in the future.  If anything, preventing future LLM from digesting output from previous AI LLMs is a positive towards better performing LLMs.

 

I'm starting to see a bit more articles from the main stream media start questioning these second order effects, so there seems to be some visible light at the end of this SaaSpocalypse.

[bargainman]

Jensen just said something like this in an interview.

 

"Imagine when we get androids in the kitchen.  They aren't going to build a cuisinart, they'll just it.  Same with a toaster, they aren't going to build a new toaster.  So the software today are the tools.  Software companies are exposing them to the LLMs which will use them to do things and as such the software will get used even more not less."  

 

S/w companies are already exposing their stuff to LLMS vs MCP and others.  APIs exist and are the foundation but you need more than that for the LLMs to talk directly to them, but not that much.  I've pointed Cursor ai at online api docs and had it build out an app in minutes to communicate to the api.  Adding it to an MCP or agentic flow is a bit more work but everyone's doing it.

[Saluki]

This is giving me bad vibes like during the dot-com and telecom bust. In that era companies would add ".com" to their name and their stock would pop. And the telecom overbuild left a lot of dark fiber, wiped out equity holders, and pissed off bondholders. 

 

Seems like people are throwing money at really crappy AI companies because they couldn't get in on OpenAI, so they want "the next ChaptGPT".  Seems like a lot of non-software companies are also touting their efforts to use AI to improve their business too. This company runs the eBay of guns, it's a very simple website that looks advanced if you were looking at it in 2000, but there is nothing about it that screams out for needing AI to make it better. But if the internet investors want AI, give them AI:  Outdoor Holding Company Launches AI-Powered Listing Tool on GunBroker Marketplace

[Castanza]

22 hours ago, elliott said:

 

Correct, SaaS companies are building agents and they will certainly fight.

But I can tell you that data from the applications is already being extracted...

Where I work, for example, is a large international company, and due to inorganic growth, there is a proliferation of applications, for example, multiple SAP systems. Data from all those SAP systems is being extracted every single day (MARA, MARC... all those tables), then integrated into a single respository, which many teams downstream are using for their own purposes. Agents that read this data are also being built and connected to the company's LLM. I don't know how it works, really, but with our LLM (just as when you use gemini or chatgpt) we can already ask questions related to our business data. E.g. what is the status of the item 00010 of order 1000056548?

 

Again, this is not the same as saying that SAP, or other SaaS, will be phased out easiliy. To begin with, the company where I work is not doing all this to phase out SAP, by any means. But you can see that things are evolving, and the future may not look like the present.

 

Yeah but permission is still being granted by your company no? Not any different that a company utilizing Splunk or some other internal tool. The gap is the public vs private data extraction and how much access new market entrants will be granted. Data ownership is still the keystone to the future. It's like owning a Ferrari and not having access to a local race track to really open the throttle up. I think encumbant SaaS companies will still have the advantage here since they already own the relationships. No wise company is going to open up their company proprietary data to some 3rd party unproven AI agent that could expose their data to their competitors.  

[elliott]

3 hours ago, Castanza said:

 

Yeah but permission is still being granted by your company no? Not any different that a company utilizing Splunk or some other internal tool. The gap is the public vs private data extraction and how much access new market entrants will be granted. Data ownership is still the keystone to the future. It's like owning a Ferrari and not having access to a local race track to really open the throttle up. I think encumbant SaaS companies will still have the advantage here since they already own the relationships. No wise company is going to open up their company proprietary data to some 3rd party unproven AI agent that could expose their data to their competitors.  

 

Correct. In our case, all of this is internal and we still have tight controls as to the data each consumer has access to, end-to-end, i.e. controls are in place in the applications that produce the data, in any of consuming application, report, or algorithm, and obviously also in the central data repository.

I dont know of us sharing any of our data with any external party (beyond what is needed for doing business with our suppliers and customers).

 

In any case, I think you are very right as to data being key.

 

 

[elliott]

11 minutes ago, elliott said:

 

Correct. In our case, all of this is internal and we still have tight controls as to the data each consumer has access to, end-to-end, i.e. controls are in place in the applications that produce the data, in any of consuming application, report, or algorithm, and obviously also in the central data repository.

I dont know of us sharing any of our data with any external party (beyond what is needed for doing business with our suppliers and customers).

 

In any case, I think you are very right as to data being key.

 

 

 

I will just add another comment in case I misunderstood the question.

 

We keep our data to ourselves, we have tight access controls, but to get the data from applications such as SAP you dont any approval from SAP itself. Many of these enterprise software need to interface with other applications, so its a feature of the software the ability to extract data, and often in a variety of ways.

[rogermunibond]

Is every legacy ERP software provider that currently runs the data layer going to have APIs and interfaces so that their agents and third-party agents can run relatively easily?  if so doesn't that move the marginal value of software from the legacy provider to the new agentic software company?  or are the legacy software providers going to create APIs that work poorly, and create walls and barriers to keep out third-party agents?

[gfp]

I thought this company's work was sort of interesting.  Taalas is the name.  They basically are turning finished AI models into a piece of hardware, all hard encoded, where it can run inference at ridiculously quick speeds and very cheaply.  I don't know if this will go anywhere but they have a model (I believe its one of the decent open source models) encoded on a large chip and have put up a demo site where you can try it out as a chatbot.

 

It is definitely not the most impressive chat bot - but you will not believe the speed at which it spits out its answer if you are used to Gemini Pro thinking or another similar "thinking" models.  Obviously there are fast software models that are optimized for speed and cost but this demo model is really really fast.

 

Company site here: https://taalas.com

 

This is the chatbot demo ->  The answers come in milliseconds  https://chatjimmy.ai

[bargainman]

2 hours ago, gfp said:

I thought this company's work was sort of interesting.  Taalas is the name.  They basically are turning finished AI models into a piece of hardware, all hard encoded, where it can run inference at ridiculously quick speeds and very cheaply.  I don't know if this will go anywhere but they have a model (I believe its one of the decent open source models) encoded on a large chip and have put up a demo site where you can try it out as a chatbot.

 

It is definitely not the most impressive chat bot - but you will not believe the speed at which it spits out its answer if you are used to Gemini Pro thinking or another similar "thinking" models.  Obviously there are fast software models that are optimized for speed and cost but this demo model is really really fast.

 

Company site here: https://taalas.com

 

This is the chatbot demo ->  The answers come in milliseconds  https://chatjimmy.ai

 

Interesting.  I wonder how they'd deal with models changing so quickly, hopefully they'd be able to update their weights.  I tried it and it was fast but didn't give me very good answers compared to Gemini Pro recently.  Could be lots of reasons for that.  I poked around a bit and it claimed it was BERT and llama based and had only 1.5billion parameters.  So it's unlikely that's the actual 30+B model.  But these days even 30+ is 'tiny'.  It also gave me very unsubstantial answers about LLM architecture so, yeah small model.

[Spekulatius]

18 hours ago, bargainman said:

 

Interesting.  I wonder how they'd deal with models changing so quickly, hopefully they'd be able to update their weights.  I tried it and it was fast but didn't give me very good answers compared to Gemini Pro recently.  Could be lots of reasons for that.  I poked around a bit and it claimed it was BERT and llama based and had only 1.5billion parameters.  So it's unlikely that's the actual 30+B model.  But these days even 30+ is 'tiny'.  It also gave me very unsubstantial answers about LLM architecture so, yeah small model.

Well if it’s hardcoded then a new model won’t run on this machine. Maybe they can do something with FPGA’s. Those are expensive however and that defies the purpose.

[bargainman]

5 hours ago, Spekulatius said:

Well if it’s hardcoded then a new model won’t run on this machine. Maybe they can do something with FPGA’s. Those are expensive however and that defies the purpose.

 

Well there are parts that could be hardcoded and other parts that could be read from memory, it's more about the architecture of the LLM and the subsequent hardware.  Beyond the one time loading of weights I doubt having that hardcoded would add much time.  It's the calcs (prefill and generative) that takes all the time and the kv cache that takes the memory.

[MungerWunger]

 

He thinks we're far away from AGI (his definition is AI that can do anything a remote worker can do)

Edited March 4 by MungerWunger

[Minseok]

4 hours ago, MungerWunger said:

 

He thinks we're far away from AGI (his definition is AI that can do anything a remote worker can do)

I heard him speak on another podcast of the same opinions.

 

its was really helpful to understand the limits and usecases of AI models.

 

One of the conclusions I came to that this algorithm is incredibly useful for programming because much of the work is structured and documented (frameworks/source codes/documentation) and source code to train from, but not as much (yet) in other traditional disciplines. He spoke about his CFO who “uses coding to execute their job” being profoundly more productive than those who work in “analog” mode. Analog being Intuitions/ tribal knowledge/ non standardized white collar labour stuff. Majority of human expertise is inside peoples heads and not open for models to train out of.

 

This would be bullish for software based enterprises like google/MS/Amazon whi will reap incredible benefits from worker productivity alone, but not in areas like engineering/law firms/ accounting firms where much of the work is still done by humans in manual ways.

 

Example on point is in the field of mathematics. There is a langauge called LEAN which makes formal mathematical proofs be expressed in a precise language like coding). This enables LLM to be much more effective in generating and verifying proofs.

 

So I see a world where when most job becomes a programming job, we will finally see significant economic gains with AI.

Edited March 4 by Minseok

[MungerWunger]

https://www.forbes.com/sites/annatong/2026/03/05/cursor-goes-to-war-for-ai-coding-dominance/

 

"Cost remains an ever present challenge. Cursor’s larger rivals are willing to subsidize aggressively. According to a person familiar with the company’s internal analysis, Cursor estimated last year that a $200-per-month Claude Code subscription could use up to $2,000 in compute, suggesting significant subsidization by Anthropic. Today, that subsidization appears to be even more aggressive, with that $200 plan able to consume about $5,000 in compute, according to a different person who has seen analyses on the company’s compute spend patterns."

 

What would happen to usage if they stopped subsidizing their users

[Minseok]

2 hours ago, MungerWunger said:

https://www.forbes.com/sites/annatong/2026/03/05/cursor-goes-to-war-for-ai-coding-dominance/

 

"Cost remains an ever present challenge. Cursor’s larger rivals are willing to subsidize aggressively. According to a person familiar with the company’s internal analysis, Cursor estimated last year that a $200-per-month Claude Code subscription could use up to $2,000 in compute, suggesting significant subsidization by Anthropic. Today, that subsidization appears to be even more aggressive, with that $200 plan able to consume about $5,000 in compute, according to a different person who has seen analyses on the company’s compute spend patterns."

 

What would happen to usage if they stopped subsidizing their users

Whats more interesting is for a company like google, where tokens are generated using internal hardware, a $5000 per head per month cost is $60,000 per year per employee. If you can make a $300,000 programmer just 20% more efficient, youre breaking even. This is a very good internal ROIC for google.

Edited March 7 by Minseok

[Spekulatius]

10 hours ago, MungerWunger said:

https://www.forbes.com/sites/annatong/2026/03/05/cursor-goes-to-war-for-ai-coding-dominance/

 

"Cost remains an ever present challenge. Cursor’s larger rivals are willing to subsidize aggressively. According to a person familiar with the company’s internal analysis, Cursor estimated last year that a $200-per-month Claude Code subscription could use up to $2,000 in compute, suggesting significant subsidization by Anthropic. Today, that subsidization appears to be even more aggressive, with that $200 plan able to consume about $5,000 in compute, according to a different person who has seen analyses on the company’s compute spend patterns."

 

What would happen to usage if they stopped subsidizing their users

They would go to cheaper more limited models (Deepseek), some of which can be run locally or on premise.

[Spekulatius]

I like how you can update your knowledge quickly by creating AI summaries . She I look or a sector or stock, I often let the Ai do some legwork and do comparison with competitors . For example, I wanted  to take a looks at CRBG (pension insurer ) and  you can create report quickly. I also know that I can just ask Gemini to create the output as a pdf document. It’s not perfect no substitution for your own  work, but a big timesaver if you want to just get started.CRBG_Equity_Analysis_March_2026.pdf

 

Here is what I came up with researching CRBG:

Edited March 7 by Spekulatius

[lnofeisone]

I didn't know where to drop this, but felt it was relevant here. This dude built an agentic system that gives live updates. You won't find a link to this specific view. Also, judging by the design, his data consumption and processing bills are likely astronomical. 

 

You can, however, take his code (there is a GitHub repo) and build something like this https://finance.worldmonitor.app/

Edited March 9 by lnofeisone

[Dalal.Holdings]

I've long been an AI skeptic, but my sentiment has been changing. AI was apparently a big part in the Maduro raid and the Iran operation... 

 

https://www.thetimes.com/world/middle-east/article/palantir-ai-software-us-iran-war-lwld892z9

 

Quote

During the US-led invasion of Iraq, a 2,000-strong American intelligence unit was asked to identify targets on the ground for western troops.

In Operation Epic Fury against Iran, 20 troops can carry out the same workload. 

They are using software developed by the CIA-backed technology company Palantir, which uses artificial intelligence to choose and dismiss targets more quickly than the human brain can think.

 

It seems it is clear that AI is integral to geopolitical strength and overall national security. And both China and the U.S. realize that. So you now have something akin to the Soviet-U.S. space race where there needs to be large amounts spent on R&D, capital intensive production, etc and the fruits of that labor translate into overall military strength. It becomes a race that is integral to the edge and overall survival odds of competing superpowers, with financial backing of governments.

 

This makes it pretty different from the tech bubble of 2000. In this case, it is becoming a national strategic imperative to lead in these technologies in China and the U.S...

 

Meanwhile, there are other places like Europe where they have pretty much thrown in the towel and destroyed this capability...just another demonstration of Europeans shooting themselves in their own foot.

[Buckeye]

Great point on the national security ramifications behind Ai. Apparently it was Claude that told the War-a-lago crew to bomb the all-girls school, based off of old data.  Now only if this group had a little more NI (natural intelligence) we’d all be better off. Too bad no one is working on developing NI. 

[jfan]

Family of 12-year-old Tumbler Ridge shooting victim files civil claim against OpenAI https://share.google/TfCodgtfr0iYewBuF

[NnnnotSoSmart]

Jobs of the future: 

 

[WayWardCloud]

Thanks for sharing! How is transportation so low? Waymo is coming for Uber drivers and truck drivers. Later on AI can also replace the delivery guys who have to drive AND get out of the car with a combination of self driving car + humanoid bot and replace Amazon / Doordash people... That is if the type of world those bros describe really is coming, which is still a big IF.