Deliberately employing someone who cannot speak or understand the language is the issue here really - also could seem a convenient way to justify pushing those roles onto LMs.
LMs have a lot of illimitations and feel there needs to be more breakthroughs……….still too much noise and propaganda in this space currently.
Just a couple technical observations of the LLM challenges.
Let us think of scenario where your colleague is a LLM. Every time you see him, he does not know who you are, where you are and what are you working on i.e. his context is empty.
Every task you start with him, he must read all the relevant information (including your or his name, what company you are working for, everything that is not global general knowledge) that has been carefully selected to not overflow his limited working memory. He also charges you by every word he reads or writes. BTW, it’s your task to prepare this info package for him.
You can also give him tools so that he can fetch more information or control external things. Every time he uses the tool he will charge for every word the tool might produce (for example get web page). Using tools are specially taxing for his limited working memory (million tokens/words is used up pretty fast).
To make things worse, his cognitive performance degrades fast if the working memory starts filling up, already significant even before half of the working memory is used. Also, he will struggle to do any new algorithmic reasoning if he did not know it by heart already (ICL issues).
And when you have conversation, he will charge you with all cumulated words in this session every time he opens his mouth or uses his tools.
It is like working with the memento protagonist, who knows a lot about things generally but has to be directed at every step or he will start wreaking havoc.
This is the state is the AI / LLM today. I would say it is very far away from general super intelligence, as some claim it to approach. 
Claude helped with the attack on Venezuela. Of course I don’t to what extent and likely never will. It’s not like Anthropic will tell all in a couple of years. There is nothing good from any of this for the bottom 99%. Yes it will drive you to your date and clean up your code but the ultimate cost will be massive.
I don’t think that’s really accurate even for the free to use LLMs like ChatGPT 5.3 is it? They hold all the previous conversations and will pull up all context as stored - and they never throw anything away.
The business plan for OpenAI is advertising, and for that you need stored context on who you are talking to. Any decent LLM platform in 2026 can have a lot of current context to access. A 2 million token context window is about thirty 50k word books, and that’s putting aside things like context compaction and Retrieval-Augmented Generation, and tools that bring in other info as needed. There’s challenges with LLMs for sure but forgetting stuff isn’t the worse one.
Technically LLM is stateless. I would guess ChatGPT has some kind of RAG system to pull data from past conversations, or maybe they embed the basic user attributes into the context to make it a bit personal. Still, these are scaffoldings and many ways inferior to real learning memory.
One issue with 2 million token contexts is that if you fill it even to half, everything will become very expensive. Each chat message or tool action will cost you a lot. Caching will help this a bit, but it is very easy to burn through lot of money then. Agentic system may well do tens of tool calls for every message exchange with it (tool calling etc). Then the costs will be in dollars per message.
Your CPU is technically stateless too but it gets by if we don’t count L1/2/3 caches, so we’d get down to terms of ‘what is a LLM’ just like ‘what is a computer’. To people using them they are a system of parts that includes context management in anything that is considered modern usage today I think.
Anyway, I just wanted to offer a respectful rebuttal to the statement that LLMs have no memory, when a big part of their used system design is actually to give them one to be more useful.
The token cost of the context window isn’t really the limiting factor in terms of cost - Gemini Pro with a large window costs less that Claude Opus for example. The real cost is the reasoning cycle so far, as that eats tokens for breakfast, and does better the larger the context window.
Historically with computing, where there is pressure of costs then there tends to be a lot of effort in bringing it down - I would expect improvements? There is a lot of money in the field working on it. Stuff like dynamic LoRA layers and context compaction is doing good things, as is hardware advances of bringing caches closer etc. It’s complicated stuff but I guess my main point was mainly don’t bet against today’s token cost as being the limit of LLM futures, as even if you go back just 2 years ago thing were very different.
Today Opus 4.5 or GPT 5.3 Codex can one-shot some forum software like this for about $50 USD and a few hours. That isn’t super cheap for a hobbyist but in the context of a team of people working on it for months is still compelling.
I think we both agree that LLM systems are incredibly useful and have a lot of untapped potential.
Many of the technical nuances come from different usage scenarios. I have been mostly using per token based pricing with API access. These models usually don’t have any long term memory baked in. The consumer UX is probably a bit different.
I think i am also venting some frustrations I have faced while using and developing the agents and apps. Somehow the 500k tokens feels so small at times. Systems like Claude Code really burns through the tokens, there’s so much more happening under the hood compared to traditional chat interfaces.
Mrianik Sharma, a lead safety researcher at Anthropic, has given up to pursue a study of poetry. He knows we’re screwed. He wants to spend what time is left leading a creative life that won’t accelerate humanity’s demise. I don’t have X so I’ll just paste his farewell letter as a photo.
Must be nice to have the choice.
Still, not a fun job to have, given the current political climate.
Imagine you’re responsible for trying to stop a multi billion dollar contract with the Department of War for a new massive surveillance operation on the US population for ethical reasons - you’ve got an IPO to prepare for, a mostly lawless kleptocracy to push back on, and competitors one hundred times larger than you offering gladly to do it more maliciously and cheaper instead.
AI is a tool. People use tools. How are the people doing?
AI is disruptive and dangerous and jobs will be lost. But being worried about AI when what is happening today is like shouting at clippy while your house burns down. This passivity and ability to not blame the right thing might be what allowed this to happen in the first place.
Humanity may demise, but I think AI might only be used to draft a more professional version of the goodbye letter.
I obviously disagree. But I applaud the beautifully written defense.
(And the fact that Clippy damn near did bring humanity to an end, but didn’t, doesn’t make those of us who railed against the dangers of Windows 95 wrong.)
Thanks.
For AI with me (definitely) and for carbon/climate change with you (maybe), this probably applies to both of us equally. 
It is difficult to get a man to understand something, when his salary depends on his not understanding it. - Upton Sinclair
At my age I am untouchable. My worry is for my daughter. She on the other hand is not worried in the least. That is as it should be.
I do feel as if I have about as good a feel for AI as anyone. That’s a bold statement from someone who can barely code on Scratch. But we’ll all recall the calls for doom from Musk, Zuck, Altman and Cook. In the two thousand teens each at one time or another had deep public fears about what it could do. Fabulous wealth and power has since silenced them. But they seem about as mystified about it today as they were back then. My biggest education is from Google’s Blaise Agũera y Arcas. He makes the case that current models display all the traits of true intelligence. But so do bacteria by his counting. Still, he’s a fan. No poetry classes at Oxford for him. This is our “Jesus take the wheel moment”. It’s just that the “Jesus” who’s about to do more of the driving is only hardly better understood than the real one.
It’s very much unintuitive but it’s not really a mystery. The Clark’s 3rd law applies to the field but then no-one worried when the stock market high frequency algorithms took over all the financial centers in the 90’s - people didn’t get how those worked, with nothing to do AI, and now they still don’t get how AI works either, it’s all ‘magic’.
The folklore about artificial general intelligence (AGI) is just noise for me personally. The biological comparisons of neural nets and the like is just fun commentary, trying to find grounded analogies for the math that is really hard to conceptualize. Calculus, derivatives, even linear algebra, are not commonly understood by most people that use AI, but they aren’t magic.
One question is that if AGI is reached then what then? What changes? Did we just change something or was some arbitrary benchmark reached? We had the Turing Test for a while until it was passed easily, so now we have ‘AGI - it’s intelligent’ as the benchmark. What does that really mean?
One lens to look through this and AI is if models can impersonate people using generative techniques based on trained weights then were people really that interesting in the first place? If your behavior can be this well mimicked through a series of patterns, loss functions and gradients, then is it the AI being smart or just people not being that varied in their thinking? Put another way, the language and vision models aren’t ‘smart’ but the concepts within them weren’t as impressive as we liked to think.
It sort of reminds me when you read a reporting piece in a field you know really well, in that you feel that it’s incredible how wrong and how much misunderstanding there is. It’s then you sort of get the sneaking suspicion that reporting in fields you don’t know about are also equally wrong and misunderstood perhaps. AI is like that for me, in that I’ve been around in it for 20 plus years now and each step up in the field has been sometimes unintuitive but not really mystery or Jesus driving.
There is a lot of money in it sounding scary/incredible though for sure.
Market is getting jumpy though. P/E ratios well in excess of 100, now where have we seen that before? 
This is from a couple of years ago now, but a good fundamental’s explainer of what is happening:
Transformer Explainer: LLM Transformer Model Visually Explained
Some of the 20 steps are a bit of a leap but it is still sound.
PS I don’t want folks here to think I see human intelligence as not amazingly special or unique (and probably non reproducible, e.g. see art), more that the results AI gives are explainable and not magic. We ‘encode’ how we communicate in language and we lose a lot in that enough to impersonate it sometimes.
There is going to be an enormous crash for sure. Like spectacular.
It’s what, six times the size of dotcom now? If that doesn’t give you pause…
Yeah, at least. I went through 2002 and it was sub-optimal 
, but the combo of the AI collapse followed by crypto hurtling down the same roller-coaster is going to make that seem like getting a $5 bounced cheque fee in comparison.
It’s the catching the knife problem though, e.g. ‘inverted yield curve has predicted 12 of the last 7 recessions’ etc.
EDIT:
Kinda off topic sorry, but I probably crossed that bridge already.
I don’t even think it’ll be AI to start a consolidation crash, it’ll just be AI crashes with crypto doing a ‘inflatable man at car dealership dance’ that’ll be watchable. The lack of US mid-term results and resultant fleeing of the bond market yields when the Fed head proxy declares inflation now doesn’t exist will likely light the fire first. AI could basically do nothing in the next 9 months and, um, ‘exogenous factors’ will just go for it. It’s the wild west coming up. 
At the risk of sounding like a broken record, AI in itself doesn’t scare me. AGI… well that’s another story, but far enough down the road that I probably won’t be alive to see it.
What terrifies me is the uses it can be put to. Sharma alludes to it in their resignation letter, it is putting the tools for mass destruction in the hands of amateurs. Safeguards can be overwritten and it won’t be long before you no longer need a PhD in biochemistry, or a solid understanding of chemistry to synthesise the types of WMDs that are currently the preserve of nation states.
