inference and you
walk blindly into the future with LLM output
have you ever accepted output from a chatbot without a second thought? had one answer a homework question to get it done? use it as source content for some other project/creation/art?
you might be surprised, like people that we talk with, that these chatbots can output incorrect content, and at times confidently make stuff up.
even knowing this, tech bros have managed to put LLM tech in the weirdest of places.
which has led people to request from conversations Python code, something a human would find weird when the context is your next vacation or phone bill, but a AI chatbot would gladly do the task.
a litmus test for LLM.
sans review
whether the output is of matter is really the criticality of the result.
you prompt a chatbot to generate an image for use on a blog, what it outputs only matters to you, and needs no review outside of your artistic eye. if the result was from a computer vision AI for early cancer detection, then the output is critical and needs human review.
context matters, as always. inference without review is a gamble.
how it’s used
how can gen-AI genuinely be used? we know cheating on your homework isn’t one of them, review what you submit students!
there are domains where it’s been a helpful tool.
- media generation
- text operations (review, search, etc.)
- coding
- biomedical research (protein folding)
cranium offloading
computer science has the concept of computation offloading, where another chip or platform does the work and provides the result.
but i do think, we have to think about the proliferation of AI as something on that level, it’s not just a tool, it’s not the bicycle, it’s not the calculator. it’s a thing that could potentially obliterate the way we think. devindra hardawar, twit #1094
you get better by doing. this goes for both physical and mental tasks.
anybody that’s taken on a project of length will understand this concept, as you look back on your project’s past work, it doesn’t include all the knowledge you learnt since doing that work, so you redo it, which can led to one of many redos.
be careful chasing perfection.
same, but different
people confuse machine learning (ML) tech that performs non-language actions with LLM tech and for good reason, non-language AI uses an LLM as a base to train another model. i don’t want GPT in my self-driving vehicle type of thinking, which is only half the truth.
starting with an LLM for the reason that LLMs have semantic knowledge about the world, that couldn’t be obtained through visual sensor training alone.
so for example, computer vision is not gen-AI, but can be, when done in reverse and used for driving simulation data.
solutions for both reach for the same tech.
how deep
for the all that gen-AI tech has brought, it has helped both adversary and ally.
faking both image/video/audio of people for malicious intent. finding 0day exploits and even autonmous AI hacking.
seaching for new phyiscal materials, folding proteins, and numberous other tasks that would take humans hundreds of thousands years to do. finding and patching the bugs before exploits are made.
so coming down on one side or another on if it’s good or bad is hard, because it can be used for both.
the environment
data centers need lots resources and infrastructure.
AI data centers are a horse of another colour. those require more, a lot more, about 100x more, and changes are happening so fast new builds are what is required, because of the increase in rack power requirements.
the water requirements are an issue, because access to fresh water was already a problem, but is becoming less of a requirement as cooling tech improves.
power is in short supply, and these AI data centers are hungry, forcing companies to get into energy generation business, with new builds and restarts, to power this AI era.
conclusion
this is useful tech that has been bestowed upon us.
how we use it is up to us.