There are a lot of grifters repackaging ChatGPT.My employer was sold an AI product based off of Chat GPT. I get to be part of the initial testing.
I asked the product's representatives what access the product had to encounter specific information. It has none.
I asked what data the product was trained on. It was trained on data from the internet.
Basically it is just ChatGPT being upsold to us. It has no specific use case and cannot actually do anything productive. It may make searching the internet slightly faster or more effective.
What a waste of money.
It's AI everything:
Still waiting for someone to tell me what 'Wi-Fi' stands for.To be fair, this is nothing new, its just the same old dipshit marketing flunkies that have reached their final form where they just search hashtags on social media and see what they need to slap on their products as a buzzword. Its worse now because it happens more quickly and is like industry-wide (including major companies that normally wouldn't operate like that), whereas in the past it would take a bit as one company finds success with it and others try to mimic. Actually this more reeks of it being the middle manager asswipes that have been put in charge of most places going "AI's huge" and slapping it on everything regardless of it making any sense whatsoever.
This happens constantly in audio for instance. Not that long ago, the audio industry officially recognized "Hi-Res" as an official term and every single company subsequently slapped a Hi-Res sticker on their stuff, nevermind "Hi-Res" products had existed for literally decades already. Dolby has pimped out Atmos so hard that it is in everything from like sub $50 mono speakers and earbuds all the way up. Plus they're paying record companies to force artists to make Atmos versions of their music. The industry tried the same thing with MQA (which really was just the latest in their long line of attempts at DRM). Honestly, I'm kinda disappointed that Dolby didn't call their push as "Atmos Integrated" to be able to cash in on AI. But then we'll probably have someone make an "AI" device that's just a tube pre-amp for vinyl and think they're being snarky by calling it "Analogue Intelligence".
Still waiting for someone to tell me what 'Wi-Fi' stands for.
I feel like people who are mad at AI for hallucinating and being unreliable aren't really 1:1 with tech. Its just like the internet, in fact they're all trained on the internet. You cannot take what results you get at face value, its all about your prompt and how you interpret the results.
Why would any expect an internet aggregator to be perfect? You aren't supposed to use google to get direct answers to things, its a tool for doing research.
Lately I've been making extensive use of AI in my work and its actually incredible what arguing with a bot can do for you. A bots hallucination might be worthless to one person, but absolutely inspiring to another. I've been really applying LLMs to help solve big hard problems and their ability to give me new leads is simply amazing. A lot of people are really sleeping on using LLMs as a tool, and instead they just want it to be a correct answer machine.
Not to say that idea is without merit, it would be nice if LLMs were always right. But in the same vein AI right now is like a hammer, compared to its ideal, a nail gun. Just because it takes more skill to use doesn't mean it isn't useful, nor does it mean there isn't a future where AI actually does make everyone happy. All tech needs iteration, duh.
I feel like I should also mention, that corporate controlled AI, services like chatGPT, are lame and I literally have never used one. I only run local models, and lately I've been using Llama 3.1. Which yes is made by facebook and you can kinda tell in the way it talks which is funny. But its still free, still local, and still useful.
We first study the common technique of post-train quantizing model weights, finding that the longer you train/the more data seen during pretraining, the more sensitive the model becomes to quantization at inference-time, explaining why Llama-3 may be harder to quantize. In fact, this loss degradation is roughly a power law in the token/parameter ratio seen during pretraining, so that you can predict in advance the critical data size beyond which pretraining on more data is actively harmful if you're serving a quantized model. The intuition might be that as more knowledge is compressed into weights as you train on more data, a given perturbation will damage performance more. Below is a fixed language model overtrained significantly to various data budgets up to 30B tokens, then post-train quantized afterwards. This demonstrates how more pretraining FLOPs do not always lead to better models served in production.
Click to expand...