Ask HN: What LLM and AI things will happen in the first half of 2023? | Hacker News Reader

Ask HN: What LLM and AI things will happen in the first half of 2023?

4 pointstikkun4 comments

Some examples / our predictions –

More obvious things:

* Internet enabled models (WebGPT)

* More context / bigger memory (GPT-4)

* Bigger models / more params (GPT-4)

* ChatGPT but you can talk to it in voice back and forth (this will be a cool demo, but won't be useful until we have a way to run it with low latency locally)

* MSFT will integrate ChatGPT into all MS office products (Clippy 2.0)

Less obvious things:

* Open source text models get more popular - GLM-130B which is superior to GPT-3 in zero and one shot learning

* ChatGPT competitors now that we understand how it was built. for example PaLM

* Lots of personalized generative art (apply styles to your house, your pets, your product photos)

* AI generated astroturfing

* We'll start running out of text and image data for training models at massive scale

What else?

View on HN

More obvious things:

* Internet enabled models (WebGPT)

* More context / bigger memory (GPT-4)

* Bigger models / more params (GPT-4)

* ChatGPT but you can talk to it in voice back and forth (this will be a cool demo, but won't be useful until we have a way to run it with low latency locally)

* MSFT will integrate ChatGPT into all MS office products (Clippy 2.0)

Less obvious things:

* Open source text models get more popular - GLM-130B which is superior to GPT-3 in zero and one shot learning

* ChatGPT competitors now that we understand how it was built. for example PaLM

* Lots of personalized generative art (apply styles to your house, your pets, your product photos)

* AI generated astroturfing

* We'll start running out of text and image data for training models at massive scale

Question: When will laptops be powerful to run a ChatGPT or GPT-3.5 like model, locally – can the M2 Max already do this? The latency of ChatGPT is a huge downside.

I'm not sure how much computing power is required, or how much disk space.

Probably never (doubly so if your laptop doesn't have Nvidia hardware acceleration). ChatGPT uses enormous 300+gb models that fit into VRAM partitioned across dozens of accelerator boards. Running the same program locally is not possible. There are generative models you can run locally (gpt neo 2.7b), but it's a much smaller model that's much less refined. Even still, inferencing it locally would probably be slower than ChatGPT running in the cloud.

There may come a day when someone has a heavily-pruned knowledge model available for local use, but I wouldn't count on it. We chose a future where everything runs in the cloud, and our "ownership" of technology boils down to a payment voucher saying we're entitled to it. OpenAI will never let you run it locally because then what would their value be?

Also, as a side note: I'm not sure these AI speculation threads are a good fit for HN. There have been so many of them in the past few months, it's been overwhelming and a little annoying. Nothing against you for posting it, but a lot of these threads have already been made in the past.

Humane will launch its first product(s).

Reply on news.ycombinator.com