HNHacker News
TopNewBestAskShowJobs

mediaman

6,587 karma · joined April 20, 2009

submissionscomments
mediaman··on Claude 3 model family
People here upset about refusals seem to not understand the market for AI, who the customers are, or where the money is.

The target market is large companies who will pay significant sums of money to save hundreds of millions, or billions, of dollars in labor costs by automating various business tasks.

What do these companies need? Reliable models that will provide accurate information with good guardrails.

They will not use a model that poses any risk of embarrassing them. Under no circumstances does a large multinational insurance company want the possibility that their support chatbot could write erotica for some customer with a car policy who thinks it might be funny to trick the AI.

It doesn't matter if you're "offended." You can use it, but you're not the user. Think about the people these are designed to replace: the customer service agents, the people who perform lots of emotional labor. You think their employers don't want a tightly controlled, cheerful, guardrailed human replacement?

mediaman··on Claude 3 model family
People like us are not the real users.

Corporate users of AI (and this is where the money is) do want safe models with heavy guardrails.

No corporate AI initiative is going to use an LLM that will say anything if prompted.

mediaman··on Big Post About Big Context
There is a lot of emerging evidence that ability to reason, and ability to compare within-context dramatically declines with large context.

It is unable to find edits: https://twitter.com/mengdi_en/status/1763338245696274469

It hallucinates changes badly: https://twitter.com/thomasahle/status/1763408041960231010

To be fair, finding diffs is an odd way to use LLMs, but does show the highly variable abilities of these models over bigger contexts.

In other models, reasoning performance declines with more tokens (no data yet here for Gemini 1.5 Pro): https://arxiv.org/html/2402.14848v1

Note in that last chart that significant declines in reasoning are apparent over just a few thousand tokens.

mediaman··on Our next-generation model: Gemini 1.5
Right now I just throw the different questions together in a single embedding for a given chunk, with the idea that there’s enough dimensionality to capture them all. But I haven’t tested embedding each question, matching on that vector, and then returning the corresponding chunk. That seems like it’d be worth testing out.
mediaman··on Our next-generation model: Gemini 1.5
A lot of people in RAG already do this. I do this with my product: we process each page and create lists of potential questions that the page would answer, and then embed that.

We also embed the actual text, though, because I found that only doing the questions resulted in inferior performance.

mediaman··on Our next-generation model: Gemini 1.5
Agreed with this.

The leading ability AI (in terms of cognitive power) will, generally, cost more per token than lower cognitive power AI.

That means that at a given budget you can choose more cognitive power with fewer tokens, or less cognitive power with more tokens. For most use cases, there's no real point in giving up cognitive power to include useless tokens that have no hope of helping with a given question.

So then you're back to the question of: how do we reduce the number of tokens, so that we can get higher cognitive power?

And that's the entire field of information retrieval, which is the most important part of RAG.

mediaman··on Eagle 7B: Soaring past Transformers
Knowledge transfer over generations is a function of the brain.

Other species have much more limited ability to transfer knowledge intergenerationally, and that is because the human brain's capability for symbolic language is much more advanced than other animals', who are not able to encode knowledge nearly as efficiently.

mediaman··on New embedding models and API updates
3.5 has JSON mode as of the November release, but only in the November-dated model: gpt-3.5-turbo-1106.

I found it to reliably produce JSON correctly, but I've found 3.5 to be a poor performer at things like entity extraction and following directions compared to other fast models such as claude-instant (though that does not have function calling).

mediaman··on Field experimental evidence of AI on knowledge worker productivity and quality
This study had management consultants use GPT4 on a variety of different management consultant tasks.

Some were more creative in nature, and some were more analytical. For example, ideas on marketing sneakers versus an analysis of store performance across a retailer's portfolio.

In general they found that GPT4 was helpful for creative tasks, but didn't help much (and in fact reduced quality) for analytical questions.

I think these kinds of studies are of limited use. I don't believe raw GPT4 is that helpful in the enterprise. Whether it is useful or not comes down to whether engineers can harness it within a pipeline of content and tools.

For example, when engineers create a system to summarize issues a customer has had from the CRM, that can help a customer service person be more informed. Structured semantic search on a knowledge base can help them also find the right solution to common customer problems.

McKinsey made a retrieval augmented generation system that searched all their research reports and created summaries of content they could use, so a consultant could quickly find prior work that would be relevant on a client project. If they built that correctly, I imagine that is pretty useful.

GPT4 alone will especially not be that useful for analytical work. However, developers can make it a semi-capable analyst for some use cases by connecting it to data lakes, describing schemas, and giving it tools to do analysis. Usually this is not a generalist solution, and needs to be built for each company's application.

Many of the studies so far only look at vanilla GPT4 via ChatGPT, and it seems unlikely that, if LLMs do transform the workplace, that a standalone ChatGPT is what it will look like.

mediaman··on Vanna.ai: Chat with your SQL database
RAG, at its core, is a very human way of doing research, because RAG is essentially just building a search mechanism for a reasoning engine. Much like human research.

Your boss asks you to look into something, and you do it through a combination of structured and semantic research. Perhaps you get some books that look relevant, you use search tools to find information, you use structured databases to find data. Then you synthesize it into a response that's useful to answer the question.

People say RAG is temporary, that it's just a patch until "something else" is achieved.

I don't understand what technically is being proposed.

That the weights will just learn everything it needs to know? That is an awful way of knowing things, because it is difficult to update, difficult to cite, difficult to ground, and difficult to precisely manage weights.

That the context windows will get huge so retrieval will be unnecessary? That's an argument about chunking, not retrieval. Perhaps people could put 30,000 pages of documents into the context for every question. But there will always be tradeoffs between size and quality: you could run a smarter model with smaller contexts for the same money, so why, for a given budget, would you choose to stuff a dumber model with enormous quantities of unnecessary information, when you could get a better answer from a higher intelligence using more reasonably sized retrievals at the same cost?

Likewise, RAG is not just vector DBs, but (as in this case) the use of structured queries to analyze information, the use of search mechanisms to find information in giant unstructured corpuses (i.e., the Internet, corporate intranets, etc).

Because RAG is relatively similar to the way organic intelligence conducts research, I believe RAG is here for the long haul, but its methods will advance significantly and the way it gets information will change over time. Ultimately, achieving AGI is not about developing a system that "knows everything," but a system that can reason about anything, and dismissing RAG is to confuse the two objectives.

mediaman··on Ask HN: Netflix doing mass layoffs today?
When people say "inflation is easing," they do not mean negative inflation. We have a word for that: deflation. "Inflation is easing" means the rate of inflation is declining. If they meant deflation, they would say deflation, not easing inflation.

One a new price level is hit, if it is then stabilized, the price level is not inherently good or bad. The goodness or badness of prices is based on the purchasing power that people have to buy those goods. Real wages are still up, so it's not obvious that "things are getting worse."

mediaman··on Landlords Are 'Handing Back the Keys' with Offices 23% Empty
What's the commercial office tenants' incentive to bail out CME debt funds?
mediaman··on Llamafile lets you distribute and run LLMs with a single file
Several OpenAI employees have said on Twitter that they are looking into this and developing a fix. It sounds as though it was not an intentional regression since they are implicitly acknowledging it. Could be an unintentional side effect of something else.

I'd expect we see improved behavior in the coming weeks.

mediaman··on OpenAI staff threaten to quit unless board resigns
The problem with this analysis is the premise: that it "takes time to hire someone."

This is not an interview process for hiring a junior dev at FAANG.

If you're Sam & Greg, and Satya gives you an offer to run your own operation with essentially unlimited funding and the ability to bring over your team, then you can decide immediately. There is no real lower bound of how fast it could happen.

Why would they have been able to decide so quickly? Probably because they prioritize the ability to bring over the entire team as fast as possible, and even though they could raise a lot of money in a new company, that still takes time, and they view it as critically important to hire over the new team as fast as possible (within days) that they accept whatever downsides there may be to being a subsidiary of Microsoft.

This is what happens when principles see opportunity and are unencumbered by bureaucratic checks. They can move very fast.

mediaman··on "Et Tu, Ilya?"
Yes, all the reporting pointed to his attempts to raise money externally for an AI chip company, and specifically soliciting sovereign wealth funds.

I could understand board concerns about him being spread too thin, but more chip competition would pretty clearly benefit OpenAI as a large consumer of GPU chips, and this is very different than any sort of allegation around mishandling of funds (which has been ruled out by several people anyway).

mediaman··on Microsoft Swallows OpenAI's Core Team – GPU Capacity, Incentives, IP
These are extremely safe conjectures.

84% of OpenAI employees have signed a statement that they will leave OpenAI if the board does not resign. The letter they signed states that Microsoft has made an open offer to any OpenAI employee. And Satya is almost certainly not going to spoil this moment with stingy packages. He is, after all, essentially acquiring a company that was valued at $90 billion, for nothing.

mediaman··on Made in China 2025
It sounds like you have a pretty antiquated view of manufacturing. It looks nothing like what you are calling it: "low value commodity bodies to screw things on." No offense to you in particular, but it sounds like you haven't updated your views since 1975. This is common because most manufacturing happens well outside of the lives of people in tech or service work.

Most screws these days are not screwed on by "bodies." They are screwed on by machine. Much of employment in manufacturing these days is in higher value work: operating CNC machines, maintaining robots and other machinery, engineering, and so on.

If manufacturing work is so low value and commodity, why do manufacturing jobs pay $20k more per person, on average, than service work?

Why does the manufacturing sector have annual average compensation of $70k, despite taking place in generally LCOL areas?

Do you look at a modern car and say: wow, what a low-value commodity good?

The Economist article you cite says nothing about manufacturing and does not support your point. In fact, when you look at the total output of manufacturing today versus 1947, it is about the same percentage of our GDP: even though manufacturing employment has dropped from 35% of Americans to less than 9% because of efficiency gains. The idea that manufacturing has disappeared from America is false and a myth.

I invite you to actually go tour a modern manufacturing facility with high speed robotic computer-controlled production lines. It may make you think differently about the physical comforts you depend on in your life.

mediaman··on Is AI the next crypto? Insights from HN comments
You don't see value in retrieval augmented generation? It seems like one of the major use cases in knowledge management in larger organizations that is hard to replicate without an LLM.

They also seem to work very well for summarizing large amounts of data, for automating the generation of basic legal texts, for extracting key data points from paperwork (invoices, mortgage applications, bank statements, etc).

It's useful to separate whether there is a lot of dubious hype (true of any new foundational technology) from whether useful things are being done. Both can be true at the same time. Lots of fraud and stupidity, but also lots of valuable work happening. With crypto, there was none of the latter, other than criminal applications.

The internet also attracted lots of hype and poor ROI consulting projects...but here we are.

mediaman··on Pretraining data enables narrow selection capabilities in transformer models
There have been two criticisms of this paper floating around.

1. The test mechanism is to use prediction of sinusoidal series. While it's certainly possible to train transformers on mathematical functions, it's not clear why findings from a model trained on sinusoidal functions would generalize into the domain of written human language (which is ironic, given the paper's topic).

2. Even if it were true that these models don't generalize beyond their training, large LLMs' training corpus is basically all of written human knowledge. So then the goalpost has been moved to "well, they won't push the frontier of human knowledge forward," which seems to be a much diminished claim, since the vast majority of humans are also not pushing the frontier of human knowledge forward and instead use existing human knowledge to accomplish their daily goals.

mediaman··on New models and developer products
Their system also does not seem to support techniques like hybrid search, automated cleaning/modifying of chunks prior to embedding, or the ability to access citations used, all of which are pretty important for enterprise search.

Could just mean it's coming, though.

mediaman··on Phind Model beats GPT-4 at coding, with GPT-3.5 speed and 16k context
In my own RAG implementations in the industrial sector, I've found it effective to first have the AI decide whether it needs to search at all. If it doesn't, the answers are much better.
mediaman··on Swedish union declares strike for Tesla employees
PTO is paying people to not work.

There is no financial difference between paying someone to not work one day every two weeks, and having them show up that day but not do anything (which we can agree would be unproductive).

Paying someone to stay home, and then paying someone else to cover them, reduces productivity of the factory because it requires more total labor to produce the same number of parts.

You can argue that the person will be more refreshed or whatever, which I think is probably a good argument for many types of jobs that are not manufacturing line work.

People who have engineering or creative jobs constantly think that you can get the same work out of four days as five. And for their jobs, because they have significantly decreasing marginal productivity per hour worked, they might be right. But that's very different than line work. The 40th hour worked does not have appreciably different productivity than the 35th hour worked in a week on a line.

mediaman··on Swedish union declares strike for Tesla employees
That lowers productivity, which is the topic in question.
mediaman··on Swedish union declares strike for Tesla employees
In manufacturing? What do you think happens when an employee doesn’t show up to the line because of high PTO or low working hours? Magic fairies run the assembly process for free?

Manufacturing does not work like software engineering. There is an extremely direct link between people showing up to the job and parts rolling off the line. The workers aren’t watching YouTube for half the day.

mediaman··on PC shipments decline slows in Q3 2023, but Apple plunges over 23%
Not the same thing.

The derivative of PC sales is increasing, from a negative number to a less negative number.

It's still shrinking, but shrinking at a lower rate.

The headline is as simple as it needs to be to be accurate.

mediaman··on EMPShield investing $1.9B with CHIPS support to build factory in Kansas
I can see how a device could protect from an EMP surge created by a grid connection. Is that what they're selling?

But that wouldn't work to protect electronics from magnetic pulses that would induce current within an electronic device's circuits. That needs grounded shielding physically around the circuit, like a Faraday cage.

mediaman··on Translating Latin demonology manuals with GPT-4 and Claude
That’s a completely different test. You’re using the vision multimodal ability to decipher Chinese script, essentially adding an OCR step to the process, and it’s not good at OCR of Chinese script.

Try feeding it actual Chinese characters. From what I understand, it’s somewhat competent.

mediaman··on Raspberry Pi 5
You are right that this metric excludes things like capital gains income, and covers just hourly earnings.

“Real”, however, refers to it being inflation adjusted, not an exclusion of high earners (who are excluded anyway because of the hourly earnings metric).

mediaman··on First Impressions with GPT-4V(ision)
No, I'm not Jakob Nielsen.

I do however run a company that employs lots of blue collar, non-college-educated people, in manufacturing. And although this is in no way scientific, my experience matches this: most people are much more uncomfortable writing than they are reading. Even with reading, most strongly resist reading documentation unless they have to, and prefer trial and erroring their own gut instinct until they happen to find something that works or they give up. (This is less true of the most highly skilled technicians, such as those who troubleshoot robots and low voltage control systems.) The official statistics on literacy are absolutely not a good indicator of how comfortable people are articulating themselves with the written word, much less reading.

This is generally met with disbelief by most people in tech I talk with about this, because for the most part they have nearly zero interaction with this large portion of the population. From their daily experience, 98%+ people can make effective use of these tools.

But almost nobody in this partially literate population wants to write in an empty text box to ask an AI to do things. They can learn to visually navigate a simple UI, especially if it's well-designed, because they can effectively make decisions about what of several paths to take.

Some others here have brought up voice, and I do agree that voice is a more promising avenue, although I think it'll still take carefully constructed conversational experiences to work well (i.e., free form 'tell it what you want' will still not work).

mediaman··on First Impressions with GPT-4V(ision)
Roughly half of people in most developed countries are not functionally articulate: meaning, they can read functionally, but struggle to articulate what they want with the written word.

LLM-based chatbots can be extremely attractive to the top 30% literacy users in the developed world. They are not a good universal UI. You still need to provide pathways for the user to follow to get done what they need without forcing them to articulate their requirement.

This is why so many people sit in front of a ChatGPT-like service and say, "what would I use this for?" and never use it again.

← PreviousPage 10 of 34Next →