Language Is Not All You Need: Aligning Perception with Language Models

The technology has been a while coming .. language models have long been a research area within machine learning, with recurrent models such as RNNs and LSTMs being an earlier approach since they allow the model to process a (language) sequence of arbitrary length.

Problems/limitations of recurrent models led to other approaches being tried using "attention" as way to let earlier parts of a sequence impact future prediction, culminating in the 2017 "Attention is all you need" paper which introduced the "Transformer" architecture that all these current LLMs are based on.

From there it was a matter of scale - scaling up the model and amount of data the models were trained on. Nobody knew how well this "Transformer" architecture could perform at scale, but early signs were promising enough to keep pushing to see how much better they could get. OpenAI in particular have been very aggressive in pushing this scaling up with their GPT-N (N=1/2/3..) models. They themselves expressed some surprise at the capabilities of GPT-2, leading to the much larger GPT-3 that is the basis of ChatGPT.

Both OpenAI and others had been leery of publically releasing these very capable LLMs for fear of ways they might be misused, but finally OpenAI released GPT-3 (with a bit of human feedback polish) in the guise of the chat bot ChatGPT, which was the first time the public had seen what the tech was capable of.

The sudden impact of ChatGPT belies the incremental improvements that brought us to this point, but seems to have been largely because the public had never seen/experienced the steps that got us here, partly because of the highly accessible packaging of the tech as a web-based chat bot, and perhaps partly because it was released without much explanation from OpenAI as to what it was/how it works - they seem quite happy for the public to do what they've done and anthromorphise it as being an AI assistant.

It's not exactly clear from the paper how they've set up the training, but it appears this model has an aspect which uses a secondary model to represent images as vectors, combines them with their text captions, and then uses those text representations along with the image vectors to train the LLM. I will leave aside the question of whether a 1024-dimensional image vector and its text caption are "images".

What's interesting is that it seems to actually lose information, as asking it to identify the studio that made WALL-E is beyond its capabilities, while asking it to describe the image (i.e. regenerating more closely something that was fed into it) and then processing on that text, is successful.

The "chain-of-thought" trick in LLMs I suspect underestimates the extent to which the interviewer is carrying water for the LLM's "reasoning" ability. The interviewer has a sense of what answer they want and will ask questions that produce further results that more easily prime the model to produce it. Reasoning supposes that these steps are carried out internally, but we see claims being made of reasoning when there is an external intelligence essentially directing the generation and combination of facts.

Another curious aspect is the flattening of 2D IQ test questions into linear format, which of course misses the point of the question in being able to reason spatially instead of linearly.

So maybe RWKV [1] is the next step. It parallelizes even better and seems to have no sequence limit.

[1] https://github.com/BlinkDL/RWKV-LM