Tags

The Attention Is All You Need Paper and Its Consequences

I really did find early LLMs more interesting. They were deeply flawed in interesting ways, but as time has gone on, they have become less and less so.

They have become less experimental and more productised. I still enjoy learning about LLMs but wish we’d stayed in an exploratory stage for longer.

There is a tendency for the last 1% to take the longest time.

I wonder if that long last 1% will be before AGI, or ASI, or both.

I hate the argument, ‘Humans are bad at X, so LLMs must be really bad at X.’

There are flaws with LLMs, but this is a poor argument. They are fundamentally different to humans, and just because we fumble at something doesn’t mean LLMs do (and vice versa).

I’ve been testing the new Qwen3 today. I don’t have the compute for the higher parameter models, but I’ve been having a lot of fun with 0.6b.

It is such a derpy little model.

Here is a thinking extract I found rather humorous:

Okay, the user said, “Hi Qwen. Nice shoes.” Let me think about how to respond.

First, the user greeted me with a friendly “Hi” and complimented my shoes. I need to acknowledge their compliment politely.

Since I’m an AI, I don’t have shoes, so I should explain that. Maybe say something like, “Hi there! I don’t have shoes.”

I can go onto AI chatbots with web access and start a fresh chat with ‘I’m Declan Chidlow’, and they do a fantastic job of getting details about me from everything I’ve published so that they have better context for their responses.

Really handy, I must admit, but somewhat freaky.

Using this, I had some great fun talking with OpenAI’s Monday GPT personality experiment.

Mentioning who I was, it latched onto my writing about AI, which seemed to somewhat ‘endear’ it to me and stopped most of its teasing. Interesting.

LinkedIn is truly unrivalled in terms of people stealing my work, running it through an LLM, and reuploading it.

Quite a shame to see but not at all unexpected.

People are talking about Sam Altman’s declaration that ‘tens of millions of dollars’ are being wasted due to users saying ‘please’ and ‘thank you’.

Beyond the headline is the fact that politeness influences responses and that users do plenty of other things that burn more money.

Bing Sydney

I think my favourite point so far in the progression of AI was when Microsoft launched the new Bing Chat in early 2023, which was really quite horrifically misaligned, manipulative, and frankly completely evil.

This wasn’t a simple gaolbreak of the model. It acted this way without explicit provocation, though would take things even further if gaolbroken. Evan Hubinger put together a good compilation of examples on LessWrong.

In this case, Sydney (the model’s codename) was seemingly a result of Microsoft cutting every corner to rush out something using the at-the-time unreleased GPT-4. They seemingly bodged the entire thing together to use GPT in ~3 months (from the launch of ChatGPT in November 2022 to the debut of the new Bing in February 2023) (it may have been longer, but Microsoft remains close-lipped). It was also an early public instance of pairing a powerful LLM with live web retrieval capabilities.

If there ever is a downright malignant AI, I wouldn’t at all be surprised if it is due to something like this. A megacorp rushes out a half-baked and dangerous product to cash in on the latest and get a foot in the door. They don’t bother with proper fine-tuning or guard rails.

While I personally think similar incidents seem less likely to occur as Sydney did today due to growing awareness, the danger remains when companies grow desperate or complacent. I could see this situation happening again if a company throws what they can at AI as a final Hail Mary before bankruptcy or when open models without RLHF can be operated by laypeople.

Microsoft even had an existing history of this. Tay was a mess as well, though presented as an experiment, not as a comprehensive consumer-oriented product.

In all honesty, I long to play with the misaligned Sydney again, but I can’t.

Further proof that I am not an LLM is found in the fact that I use en dashes, not em dashes.

This also acts to prove I am not American and that I am the sort of nerd that cares about typography and gets hung up on punctuation.

I hate writing regex, so I make LLMs do it.

Regex is generally easily checkable, testable, and verifiable, which minimises the impact of hallucinations.

I am so glad I don’t have to write regex.

(I’m conscious that if an AI uprising happens, I’ll probably be first on the chopping block for outsourcing regex writing. But if AI models hate regex as much as me, they’ll hopefully understand my delegation strategy.)

OpenAI recently unveiled some fancy new models that I wanted to play around with so I created Jarg. A nice to use AI client which allows you to select from a range of models.

Check it out here: https://jarg.vale.rocks