Tools are wielded by the mind, not ultimately the hand. The hand itself is a (flexible) tool of the mind

LLMs glom onto pre-existing tropes, idioms, and verbal constructs.

And then subtly misuse them.

LLMs write like some people cook, adding bespoke artisanal Belgian sea salt, truffle oil and weirdly specific cheese without any thought given to how the result will taste.

> And then subtly misuse them.

For example?

No, your constant output example isn’t what people are talking about with “hallucination.” It’s not about destroying information from the input, in the sense that it you asked me a question and I just ignored you, I’m not in general hallucinating. Hallucinating is more about sampling from a distribution which extends beyond what is factually true or actually exists, such as citing a non-existent paper, or inventing a historical figure.

Sure, but if we label the text as “factually accurate” or “logically sound” (or “unsound”) etc., then we can presumably greatly increase the probability of producing text with targeted properties

A pixel (in the context of an image) could be “wrong” in the sense that its assigned value could lead to an image that just looks like a bunch of noise. For instance, suppose we set every pixel in an image to some random value. The resulting would look like total noise and we humans wouldn’t recognized it as a sensible image. By providing a corpus of accepted images, we can train a model to generate images (arrays of pixels) which look like images and not, say, random noise. Now it could still generate an image of some place or person that doesn’t actually exist, so in that sense the pixels are collectively lying to you.

My recollection of CLIP is that it’s more of a text-language co-embedding, where you have two transformers, one which encodes images into vectors and one which encodes captions into vectors. Through a contrastive loss (positive pairs are captioned image pairs, negative pairs are random image-caption combinations), embeddings of positive image-caption pairs are brought together (i.e., made similar) and embeddings of negative image-caption pairs are made more dissimilar.

https://openai.com/index/clip/

I would assume some of them can write Verilog code

Well, I clearly had to try it even though I don't know my verilog from my veri-println. Can someone who does know, rank this?

--

Prompt: verilog code for 4 bit binary adder

[I'm skipping the English description of the code, this is just the code part of free ChatGPT's response, no cherry picking]

  module binary_adder(A, B, Cin, S, Cout);
    input [3:0] A, B;
    input Cin;
    output [3:0] S;
    output Cout;
    
    wire C0, C1, C2;
    
    // Full adder for bit 0
    full_adder FA0(A[0], B[0], Cin, S[0], C0);
    
    // Full adder for bit 1
    full_adder FA1(A[1], B[1], C0, S[1], C1);
    
    // Full adder for bit 2
    full_adder FA2(A[2], B[2], C1, S[2], C2);
    
    // Full adder for bit 3
    full_adder FA3(A[3], B[3], C2, S[3], Cout);
    
  endmodule
  
  module full_adder(A, B, Cin, S, Cout);
    input A, B, Cin;
    output S, Cout;
    
    assign S = A ^ B ^ Cin;
    assign Cout = (A & B) | (A & Cin) | (B & Cin);
    
  endmodule

It seems correct, but verbose. GPT-4 gives the more concise and idiomatic:

  module full_adder(input [3:0] a, input [3:0] b, input cin, output reg [3:0] sum, output reg cout);

  always @(*) begin
      {cout, sum} = a + b + cin;
  end

  endmodule

We found that Codex was pretty bad at Verilog and so we fine-tuned our own models on Verilog code, starting from the SalesForce CodeGen model. The paper will be at DATE: https://arxiv.org/abs/2212.11140

The largest model we trained is here: https://huggingface.co/shailja/fine-tuned-codegen-16B-Verilo...

Data can be crowd sourced, too. Wikipedia demonstrated that crowdsourced data can be pretty competitive.

More recently the open LAION data sets have become widely used by both tech giants and independent researchers.

> Wikipedia demonstrated that crowdsourced data can be pretty competitive.

The problem is DL is really sensitive to dirty data, disproportionately so.

At $DAYJOB once we cleaned the dataset, removed a few mislabeled identity/face pairs (very few, about 1 in 1e4) and the metrics goes up a lot.

You need to be very careful about making sweeping generalizations based on a single personal anecdote. The really large data sets typically have very high error rates and sample biases. For instance, Google’s JTF300M is far noisier than ImageNet, which itself is hardly free of errors and biases. Any data set with hundreds of millions to billions of images will generally contain a large proportion of images and labels scraped from the web, w/ automatic filtering or pseudolabeling, perhaps w/ some degree of sampled verification by human labelers.

In fact, generally DL is quite tolerant to label noise, especially using modern training methods such as SSL pretraining.

https://arxiv.org/pdf/1705.10694.pdf https://proceedings.neurips.cc/paper/2018/file/a19744e268754... https://proceedings.mlr.press/v97/hendrycks19a.html

If it faithfully memorized and reproduced a set of watermarks, it would be premature to conclude that it hadn’t memorized other (non-generic) graphical elements.

Yes, the future is now. It’s still early, extracted model quality will undoubtedly improve dramatically over time.

https://blogs.nvidia.com/blog/2022/09/23/3d-generative-ai-re...

https://research.nvidia.com/publication/2021-11_extracting-t...