AI news recap for July: While Hollywood strikes, is ChatGPT getting worse?

Hollywood actors strike over use of AI in films and other issues

Synthetic intelligence can now create pictures, novels and supply code from scratch. Besides it isn’t actually from scratch, as a result of an enormous quantity of human-generated examples are wanted to coach these AI fashions – one thing that has angered artists, programmers and writers and led to a collection of lawsuits.

Hollywood actors are the most recent group of creatives to show towards AI. They worry that movie studios might take management of their likeness and have them “star” in movies with out ever being on set, maybe taking over roles they’d quite keep away from and uttering traces or appearing out scenes they’d discover distasteful. Worse nonetheless, they may not receives a commission for it.

That’s the reason the Display Actors Guild and the American Federation of Tv and Radio Artists (SAG-AFTRA) – which has 160,000 members – is on strike until it can negotiate AI rights with the studios.

On the identical time, Netflix has come under fire from actors for a job itemizing for individuals with expertise in AI, paying a wage as much as $900,000.

AIs trained on AI-generated images produce glitches and blurs

Talking of coaching information, we wrote final yr that the proliferation of AI-generated pictures could possibly be an issue in the event that they ended up on-line in nice numbers, as new AI fashions would hoover them as much as practice on. Consultants warned that the end result would be worsening quality. On the danger of constructing an outdated reference, AI would slowly destroy itself, like a degraded photocopy of a photocopy of a photocopy.

Nicely, fast-forward a yr and that appears to be exactly what is occurring, main one other group of researchers to make the identical warning. A crew at Rice College in Texas discovered proof that AI-generated pictures making their method into coaching information in giant numbers slowly distorted the output. However there’s hope: the researchers found that if the quantity of these pictures was stored beneath a sure degree, then this degradation could be staved off.

Is ChatGPT getting worse at maths problems?

Corrupted coaching information is only one method that AI can begin to disintegrate. One research this month claimed that ChatGPT was getting worse at arithmetic issues. When requested to examine if 500 numbers have been prime, the model of GPT-4 launched in March scored 98 per cent accuracy, however a model launched in June scored simply 2.4 per cent. Unusually, by comparability, GPT-3.5’s accuracy appeared to leap from simply 7.4 per cent in March to nearly 87 per cent in June.

Arvind Narayanan at Princeton College, who in one other research discovered different altering efficiency ranges, places the issue right down to “an unintended aspect impact of fine-tuning”. Principally, the creators of those fashions are tweaking them to make the outputs extra dependable, correct or – probably – much less computationally intensive to be able to reduce prices. And though this may increasingly enhance some issues, different duties may undergo. The upshot is that, whereas AI may do one thing effectively now, a future model may carry out considerably worse, and it is probably not apparent why.

Using bigger AI training data sets may produce more racist results

It’s an open secret that numerous the advances in AI in recent times have simply come from scale: bigger fashions, extra coaching information and extra pc energy. This has made AIs costly, unwieldy and hungry for sources, however has additionally made them way more succesful.

Actually, there’s numerous analysis happening to shrink AI fashions and make them extra environment friendly, in addition to work on extra sleek strategies to advance the sector. However scale has been an enormous a part of the sport.

Now although, there’s proof that this might have severe downsides, together with making models even more racist. Researchers ran experiments on two open-source information units: one contained 400 million samples and the opposite had 2 billion. They discovered that fashions educated on the bigger information set have been greater than twice as more likely to affiliate Black feminine faces with a “felony” class and 5 instances extra more likely to affiliate Black male faces with being “felony”.

Drones with AI targeting system claimed to be ‘better than human’

Earlier this yr we coated the unusual story of the AI-powered drone that “killed” its operator to get to its supposed goal, which was full nonsense. The story was quickly denied by the US Air Force, which did little to cease it being reported all over the world regardless.

Now, we’ve got contemporary claims that AI fashions can do a greater job of figuring out targets than people – though the details are too secretive to reveal, and due to this fact confirm.

“It could examine whether or not persons are sporting a specific sort of uniform, if they’re carrying weapons and whether or not they’re giving indicators of surrendering,” says a spokesperson for the corporate behind the software program. Let’s hope they’re proper and that AI could make a greater job of waging battle than it could possibly figuring out prime numbers.

When you loved this AI information recap, attempt our particular collection the place we discover probably the most urgent questions on synthetic intelligence. Discover all of them right here:

How does ChatGPT work? | What generative AI really means for the economy | The real risks posed by AI | How to use AI to make your life simpler | The scientific challenges AI is helping to crack | Can AI ever become conscious?