The secret to Sparrow, DeepMind's latest Q&A chatbot: Human feedback

DeepMind has trained a chatbot named Sparrow to be less toxic and more accurate than other systems, by using a mix of human feedback and Google search suggestions. Chatbots are typically powered by large language models (LLMs) trained on text scraped from the internet. These models are capable of generating paragraphs of prose …

[M.V. Lipvig]

"Sparrow is a research model and proof of concept, designed with the goal of training dialogue agents to be more helpful, correct, and harmless. By learning these qualities in a general dialogue setting, Sparrow advances our understanding of how we can train artificial general intelligence to be cheaper and more useful – and ultimately, to help replace human agents," DeepMind explained.

FTFY.

[Snowy]

Human speach

So easy even a child can do it.

[amanfromMars 1]

Re: Human speach

Nice booby trap, Snowy. :-)

[Jan K.]

I've said it before... it's an insult to everything living using the word "intelligence" when talking of these cut'n'paste algorithmes.

It has nothing even remotely to do with intelligence.

[FeepingCreature]

I think it does.

[amanfromMars 1]

Being agreeable

So do I, Feeping Creature, thinking/imagining the programming revolving around little else.

[CrackedNoggin]

So much human output is simply cut'n'paste having nothing nothing remotely to do with anything but transparent pattern matching - yet it is "human intelligence". The bar is pretty low.

[Pascal Monett]

"Chinchilla has 70 billion parameters"

And I'm sure that no human has controlled more than a thousand of them.

Honestly, parameter count for pseudo-AI is the exact same thing as pixel count for screens. The hyperbole is unending.

[Anonymous Coward]

Re: parameter count ... is the exact same thing as pixel count for screens

It seems to me that parameter count is a reasonable measure of *size*, which, in the absence of other measures, is not entirely useless. But of course "size" is not the same thing as a measure of utility or performance.

[Ken Moorhouse]

Re: But of course "size" is not the same thing as a measure of utility or performance.

Many websites have researched this...

(NSFW)

[Anonymous Coward]

I think such AI could be trained to be the next British monarch, at a fraction of the cost of the current royal family.

[Anonymous Coward]

Better than a lot of humans I know.

...capable of generating paragraphs of prose that are, at a surface level at least, coherent and grammatically correct...

[Ken Moorhouse]

Sparrow

Will it be cheap?

[breakfast]

The Register has asked DeepMind for further comment

Handling media requests would be a great job for an AI chatbot and might result in some more interesting press releases when it goes off-script.

[amanfromMars 1]

Re: The Register has asked DeepMind for further comment

Handling media requests would be a great job for an AI chatbot and might result in some more interesting press releases when it goes off-script. ..... breakfast

The question to always ask of press releases whether they go off-script or not is ...... Whose script are they and its supporters promoting and prompting/aiding and abetting?

[innominatus]

Re: The Register has asked DeepMind for further comment

Shirley The Register should have asked Sparrow itself?

[Il'Geller]

NIST TREC QA

What happened to the GO game, which Brin and Page claimed was Artificial Intelligence? In the end, the couple is forced to admit that AI is the finding answers to the questions asked, in their context and subtext. That is, the solution to the challenge posed by NIST TREC QA.

[Version 1.0]

Re: NIST TREC QA

Me; "What's a one plus a one?" ... AI; "That's eleven"

Me; "No try again" ... AI; "Sorry, yes it's 10"

Me; "No, decimal, not binary" ... AI; "You are incorrect, two binary 1's create 10 binary."

Me; "Oh yes, the answer is two!" ... AI: "No you are wrong, 2 is not a binary number."

[RM Myers]

"...relevant link from Google search ... about 78 per cent of the time."

So about 7 times the percentage of relevant links that I usually get from Google search (I'm being optimistic here). Then again, they may have a looser definition of relevant.

[Evil Auditor]

...software, however, often picks up bad traits from the source material resulting in it regurgitating offensive, racist, and sexist views...

Simple criteria to detect a machine: if the answers given do not express offensive, racist, or sexist views, it will not pass the Turing test.