I have to agree with Linus tho. LLM’s are a tool, and it’s societies responsibility to use it well. Damning the tool because some people don’t wield it correctly or misuse it wont help - the genie is out of the bottle, and even tho I am very sure OpenAI and Anthropic have no future because they are hemorrhaging cash from every possible spot, LLMs and specialized models for specific use cases wont disappear.
It’s like angle grinders. They can be incredibly dangerous, as in life threatening, especially with the wrong disk or used for the wrong purpose, but in the hands of a knowledgeable person, with protection, and with the appropriate disk, they are a pretty awesome and effective tool.
Because they are great tools everywhere you can easily and by automated systems verify the correctness of their output, that’s why they are so good at e.g. finding security issues - it’s trivial to check if the LLM’s output lead to remote code execution or privilege escalation.
Testcases for software in general are something that are rather boring work - i see no issue in offloading that. And even tho the amount of Devs that document their code has increased since the 80s, i’d prefer an LLM-annotated source over one without comments at all.
The LLM doesn’t have to understand what it is doing to be effective - similar to the 10000 monkeys on typewriters writing everything that can be written given infinite time, but with a very much more educated guess than random gibberish.
They also enable natural language interfaces that are far more powerful than Alexa or Google Assistant ever were, they might be used in games for generating the lore and dialogue for all NPC in entire worlds (or even the entire world itself) by training small scale models with whatever information game writers want to provide as basis.
/// Edit here: This would also enable self-created worlds! Just create a tool to train a small model (probably available for a larger model as a tool) based on whatever lore dump you wanna provide, and then go explore that world with your own eyes!
There are many use cases for LLMs, and none of them have to be as wasteful and hostile as the crap the hyperscalers are currently pushing. Damn, if they would stop building datacenters and hoovering up every piece of silicon they can, prices for a GPU with enough VRAM to run 27b models would be trivial for the average costumer, and those can be very sophisticated if specialized. Also wouldn’t need gas turbines for power, huge complexes that have no other possible usage, no evaporative cooling.
I have to agree with Linus tho. LLM’s are a tool, and it’s societies responsibility to use it well. Damning the tool because some people don’t wield it correctly or misuse it wont help - the genie is out of the bottle, and even tho I am very sure OpenAI and Anthropic have no future because they are hemorrhaging cash from every possible spot, LLMs and specialized models for specific use cases wont disappear.
It’s like angle grinders. They can be incredibly dangerous, as in life threatening, especially with the wrong disk or used for the wrong purpose, but in the hands of a knowledgeable person, with protection, and with the appropriate disk, they are a pretty awesome and effective tool.
I agree that LLMs won’t disappear even after the bubble bursts. But I fail to see why that should be an excuse for using them.
Because they are great tools everywhere you can easily and by automated systems verify the correctness of their output, that’s why they are so good at e.g. finding security issues - it’s trivial to check if the LLM’s output lead to remote code execution or privilege escalation.
Testcases for software in general are something that are rather boring work - i see no issue in offloading that. And even tho the amount of Devs that document their code has increased since the 80s, i’d prefer an LLM-annotated source over one without comments at all.
The LLM doesn’t have to understand what it is doing to be effective - similar to the 10000 monkeys on typewriters writing everything that can be written given infinite time, but with a very much more educated guess than random gibberish.
They also enable natural language interfaces that are far more powerful than Alexa or Google Assistant ever were, they might be used in games for generating the lore and dialogue for all NPC in entire worlds (or even the entire world itself) by training small scale models with whatever information game writers want to provide as basis.
/// Edit here: This would also enable self-created worlds! Just create a tool to train a small model (probably available for a larger model as a tool) based on whatever lore dump you wanna provide, and then go explore that world with your own eyes!
There are many use cases for LLMs, and none of them have to be as wasteful and hostile as the crap the hyperscalers are currently pushing. Damn, if they would stop building datacenters and hoovering up every piece of silicon they can, prices for a GPU with enough VRAM to run 27b models would be trivial for the average costumer, and those can be very sophisticated if specialized. Also wouldn’t need gas turbines for power, huge complexes that have no other possible usage, no evaporative cooling.
Second Edit: https://www.propublica.org/article/anthropic-mythos-microsoft-software-vulnerabilities displays how awesome these things are at finding vulnerabilities. Using LLMs to find security issues should be implemented regardless of personal stance towards LLM, to make sure we all have technology we can rely upon.
What’s the problem with coding LLMs? They are insanely useful.
Maybe go after the machines that are hacking basic utilities, building biochemical weapons, doing mass surveillance, etc…