The key difference is that while an encyclopedia holds facts themselves, LLMs trained on that source material encode something more like a highly probable facsimile of those facts - the original fact was lost, LLMs are lossy, but can often be generated again with a decent level of accuracy by churning through stats about words, concepts, and relationships between them.
The whole catch is that they can often be regenerated. But LLMs (on their own, in their parametric memory - which is the result of training) don't have any conception of whether what they've generated is a real reproduction of some training material or whether they've invented something false that seemed probable based on their encoded stats. When the probability produces something contrary to what was in the training material, you get hallucinations.
They're very, very good predictive text models and can be very, very powerful when hooked up to other tools or outside databases. But its fundamentally lossy technology and all the books having been fed in doesn't guarantee all of the knowledge from those books can be spat back out.
Humans can learn texts by heart, even lots of text (I have some expertise on that, having studied opera singing). They can reproduce these texts accurately, deterministically and repeatably. An LLM is a statistical machine. It does not know any text by heart and it is by pure chance that it sometimes reproduces existing texts verbatim.
Break it into small chunks and practice. Practice more. Put the chunks together and practice even more. Like any skill, it isn't something people are magically good at beyond minor proficiency.
Every performance- singing a song, playing an instrument, performing a stand-up routine, giving a speech, performing a theatrical role are all things that are best done from memory but require practice.
There are pneumonic tricks you can use- I've seen some people do it for tricks like memorizing the order of a deck of cards- but it's less useful for long term recital because it helps with order but not comprehension or fast indexing.
Humans sometimes don't remember things 100%. LLMs can produce certain text and be run deterministically.
I don't really understand this side of the debate other than as a gotcha tbh.
If LLM use atrophies your brain and skills that's bad. If it has a repulsive writing style that's bad.
I'm not sure what the debate about whether an AI is a statistical parrot unlike humans accomplishes. Is relying 100% on a bad human speechwriter somehow better?
It comes up frequently when discussing whether LLMs actually have intelligence or qualify as AI, especially in discussions on the path to achieving AGI.
The primary complaint seems to be that LLMs are held to a higher standard than humans, though I don't particularly buy that line of reasoning.
People tend to mix the terms around a lot, but generally you'll see the manufacturers use "ducted split system" or "mini split", where mini splits are the self-contained wall or ceiling mounted units.
I live in an NYC apartment with a similar Mitsubishi ducted split system - very short run ductwork just because the apartment itself is small, but takes up way less space than traditional American-style air handlers, even the apartment sized ones. The entire air handler is suspended above my bathroom ceiling, in between the drywall and the slab above.
My system was originally installed with the 2-wire serial wall thermostat but I converted it to one of Mitsubishi's Redlink wireless units because I needed to relocate the thermostat. But since Redlink isn't wifi, and the Mitsubishi networked thermostats and cloud platform get pretty poor reviews... this project is super interesting to me.
I think there's an element of this which really breaks down to the type system being the part of formal verification that we've figured out how to do during the course of implementation.
Software engineers (myself included, over the years) often argue their real value isn't just writing code, its figuring out the gaps in requirements and how to resolve them. Sometimes that engineering process gets turned back into a formal spec. But much more often, the implementation functionally becomes the spec and contains many details that were never present in the original statement of the requirements.
Formal verification techniques in general are a harder sell until we get the industry to a point where there's broader agreement that what we call "implementation" is often a blurry mix of spec development, prototyping, and actual implementation all happening at the same time.
The details people gloss over when throwing SOC 2 or whatever other audit costs around are the complexity of the system being audited, the chosen criteria to audit (AICPA defines 5 families of criteria... Security is one, but you can optionally add Processing Integrity, Confidentiality, etc etc) and the reputation of the auditor.
A security-only audit for a small company with a narrow product focus can indeed be very inexpensive. A full SOC 2 examination for a large organization with a mix of legacy and modern systems by a name-recognizable public accounting firm can be many hundreds of thousands of dollars, or more if you need a Big 4 firm.
In my opinion, there's not much value to the "cheap" audits... If you're doing enterprise sales to a certain kind of client, your partners who demand an audit are going to want a reputable auditor or they're just going to put you through their own in-depth procurement due diligence regardless. The segment of the industry where a SOC 2 attestation is mandatory to participate but where any random auditor will do feels pretty narrow.
Unless you have a very good reason (I compare notes with people at dozens of firms and have never heard one), the only criteria you ever want to get SOC2'd on is Security.
My experience is the opposite of yours: having a security SOC2 ends the vendorsec process it any enterprise buyer, and enterprise buyers virtually never read anything in the SOC2 other than a glance at the exceptions. A very large, very security-intensive vendor we have all heard of told me a story about a vendor they had that gave them several years of repeated Type 1 reports. Went fine.
The real takeaway for projects and companies should be that someone having historically behaved in a logical and responsible way doesn’t guarantee that they’ll continue to do that for forever.
Good security architecture has circuit breakers, even for people who are generally high-trust.
I’m one of those people. Wore watches regularly through my mid 20s, completely fell out of the habit as I spent more years working from home and my routine around “getting ready for the day” loosened, and the Apple Watch was the thing that got me to put something on my wrist again - until I got sick of the screen and kept wearing watches, but now analogue ones.
The concept of "tool building" is one of the areas my team has spent the most time coaching our less-technical employees on since widespread LLM rollout in our company.
Developers and developer-adjacent, technical people tend to think this way on their own... but every business has dark corners where repetitive, manual things still happen. We're leaning a lot on training and even org-wide LLM instructions to try and let the LLM (by its own assessment) be the vehicle use to codify a process and turn it into some good old-fashioned reviewable, deterministic automation.
Honestly whipping up a lexer/parser and a REPL is one of my favorite ways to learn a new language. You can cover a lot of ground in a "real" language by just doing the frontend implementation of your own made-up language grammar and a little eval loop and its great for learning/teaching because you don't get bogged down in trying to solve some actual problem.
Which is to say: no shame in just settling for that simple C backend!
Whats implied there is that the "human touch" will become a luxury.
I have a hard time seeing a future where retail doesn't bifurcate even further into ultra low margin, happy-path optimized megastores and concierge-style high touch boutiques. Places like Crutchfield that split the difference nicely seem to be a dying breed.
Even in a relatively open organization where conversations and work are public/discoverable by default, there's still a huuuge difference between the level of curiosity required to join a convo happening in the office kitchen while you're waiting for a coffee to brew vs needing to spend your idle time at work discovering places (Slack channels or whatever else) to chime in while hoping you're not a distraction for others.
I'm a pretty staunch defender of remote work for most roles, but outside of the smallest companies where the entire organization is on a single conversational thread, you really do lose the organic peripheral vision that comes with an office environment and deliberate effort is required to try and recreate some of that in your fully-remote org if you want some of the same upside. Even with deliberate effort, I'm not convinced you can match it perfectly.
In a large company you aren’t going to be in the same country let alone kitchen. Ok can eavesdrop on conversations across our slack channels no matter where they occur, in person I’d be limited to the small subset of people I sit near.
The whole catch is that they can often be regenerated. But LLMs (on their own, in their parametric memory - which is the result of training) don't have any conception of whether what they've generated is a real reproduction of some training material or whether they've invented something false that seemed probable based on their encoded stats. When the probability produces something contrary to what was in the training material, you get hallucinations.
They're very, very good predictive text models and can be very, very powerful when hooked up to other tools or outside databases. But its fundamentally lossy technology and all the books having been fed in doesn't guarantee all of the knowledge from those books can be spat back out.
reply