dandi8

Stating you hate LLMs because you care for humanity directly supposes that those who do not hate LLMs, do not care for humanity.

You accused me of tribalism. I responded that the only tribal thing about my stance is that I care about the "tribe" of humanity. How you got what you got out of it, I truly don't know. You were the one throwing out ad hominems, dude.

So your issue is specifically with interacting with AI using natural human language?

At this point I have no idea what you're on about. The concerns regarding LLMs are widely documented. Some examples:

• They must be trained on stolen data to be vaguely useful. Yes, training on the Common Crawl still counts as stealing. No, there's not enough royalty free data to train on which would create a useful LLM.
• "Useful" in this context is extremely debatable. The architecture itself makes hallucinations a mathematical certainty. The chatbot must use the internet to have general up-to-date knowledge and that opens it up to prompt injection.
• Security is an unsolvable problem for LLMs. Even if you don't give it access to the internet, it can still be prompt injected via an external document. Prompt injection cannot be fixed because all inputs go through the same place - the prompt. 'Agentic' use cases are hilariously insecure and are, again, insecurable. A system prompt "guard rail" is not a security measure.
• Context rot is another probably unsolvable problem which renders large context windows useless.
• the amount of electricity and water needed to train the LLM and then infer outputs is untenable (there is no proof inference is profitable).
• The amount of hardware necessary to power LLMs is untenable (no, local quantized models are not "almost as good" and were not trained for free). I can't buy a new PC and that's insane.
• Because the LLMs are just fancy autocomplete (yes, they are, and no amount of backpropagation and attention heads will change that), they just produce most likely text, not actual answers. As such, the answers are often incorrect but *sound *like they are. Therefore, no LLM can be trusted to produce accurate information at any point in time.
• The entire "AI" industry is unprofitable and propped up on debt and circular financing by an industry that's out of hypergrowth ideas, while gaslighting normal people that it's the future.
• There is no future where this ends well. Either the bubble bursts and the economy collapses, or "AI" takes our jobs and we're left to starve.
• LLMs are a fun toy, but no one has found a valuable use case for it that couldn't have been achieved using other means. Oh, people are vibe coding their own software with it? Then where are all the world-changing startups that completely transform our lives? Where's a single "AI" success story that people are excited to use, besides some people falling for the sycophancy of chatbots?
• The existence od LLMs has objectively made the world a worse place, due to AI slop and misinformation. It was also used as an excuse when companies lay people off. LLMs are, terrifyingly, used in medicine, where they hallucinate patient notes saying the wrong breast has cancer, or that the patient is a drug addict (when they're not). They're being used to avoid accountability when bombing schools. CEOs and managers uncritically enforce LLM adoption despite the known and very obvious risks and limitations.

I could go on. There's so much more. No, I'm not against machine learning. I'm not against deep learning. I'm against LLMs - a technology which takes human input and outputs text and needs enormous amounts of human text, electricity and water to produce a fancy autocomplete which has extremely narrow use cases at best and is used to enshittify the entire world while stealing our resources.

But the original comment was about how you can't trust AI (in this context LLM) output. You still can't. You still shouldn't.

As far as we know, no AI generated output can ever be trusted without careful verification. Only deterministic algorithms can be given that level of trust.

But let's just say for the sake of argument that you're 100% correct, and that Evo is not an LLM.

Why only assume? I cited Wikipedia. You cited nothing.

That is why I absolutely reject your disrespectful framing that by defending a technology, and NOT it's worst uses, I'm somehow opposite to "care(ing) for humanity."

There you go putting words in my mouth again.

The actual technology behind Evo and ChatGPT is structurally the same.

...Except that it isn't. It's not a Generative Pre-Trained Transformer. It uses a Transformer-like architecture. You cannot use Evo's architecture to make a chatbot. It's a GLM.

The methods of training a model on genomic data or weather patterns is indistinguishable from training it on stolen media.

Go ahead and train a StripedHyena2 model to be a chatbot, then. I'm sure that will work great.

For someone who's such a stickler for making 100% correct and unambiguous statements, you're sure keen on asserting equality where there is merely similarity.

Nobody is claiming these technologies don't share some (or even a lot of) DNA. Being upset that people correctly use the definition of LLMs as outlined by Wikipedia, where even Evo's own Github page doesn't claim it's an LLM, is just derailing the conversation away from people's righteous objections.

Nobody is campaigning against using non-chatbot AI for science.

Read the room. Pay more attention to the context.

You don't understand what LLMs are, because LLMs ARE deep learning models, and instead of taking the time to actually learn about the technology you're responding to my corrections with hostility.

You're claiming that I don't understand technology while seemingly claiming that because LLMs are a type of deep learning, then all deep learning models are LLMs.

Evo was trained on genomic sequences, not human text. Per Wikipedia:

A large language model (LLM) is an AI model (typically a neural network) trained on a vast amount of text for natural language processing tasks, especially language generation.

Genomic sequences are not natural language. Ergo, "definitionally" Evo 2 is not an LLM.

While its StripedHyena2 architecture is very similar to LLMs, it does not use the same Generative Pretrained Transformer architecture associated with LLMs (per: https://docs.nvidia.com/bionemo-recipes/2.6.3/interactives/illustrated-evo2/index.html ).

My hate of LLMs is certainly not misinformed. It's only tribe-based in that I care for humanity.

if you think LLMs can only be chat-bots, and can only be corporate, and can only be trained unethically

Don't put words in my mouth.

1 and Evo 2 for example "Speak" genome sequences. There are other models that have been used to improve weather modeling, and animal behaviour analysis.

Evo 1 and 2 are deep learning models (I think Genomic Language Models would be the correct term), but not Large Language Models. I'm willing to bet neither are the "other" models you're mentioning. They're unlikely to be trained on vast amounts of human text for purposes of natural language interaction.

Besides, none of those "other models" are in any way applicable to either the OP or the critique of using LLMs as a source of information.

Obviously I'm using AI here as a shorthand for (all) LLMs, like the author of the article. The term AI is, on its own, so broad and meaningless, as to be entirely useless without a proper context to scope it.

No LLM (corporate or not) can be a reliable source of information due to the architectural limitations of LLMs.

Just like all those developers from the study who thought they were 19% faster but were, in fact, 10% slower?

Fair enough, it makes Jensen Huang's life better, I suppose.

Or maybe just don't use AI at all? It's not a reliable source of information, it's unsustainable in terms of resources consumed, it's unethically trained and it makes everyone's lives worse.

A slap to the face? AI isn't a reliable source of information.

And how am I going to get past the bank's custom 2FA which requires the phone app?

Per the original reply to your top comment:

Are you seriously comparing using generative diffusion models to applying a chroma key in video editing?

So you agree, then, that comparing the two is like comparing apples to oranges?