ChatGPT’s newest GPT-4 upgrade makes it smarter and more conversational

AI just keeps getting smarter: another significant upgrade has been pushed out for ChatGPT, its developer OpenAI has announced, and specifically to the GPT-4 Turbo model available to those paying for ChatGPT Plus, Team, or Enterprise.

OpenAI says ChatGPT will now be better at writing, math, logical reasoning, and coding – and it has the charts to prove it. The release is labeled with the date April 9, and it replaces the GPT-4 Turbo model that was pushed out on January 25.

Judging by the graphs provided, the biggest jumps in capabilities are in mathematics and GPQA, or Graduate-Level Google-Proof Q&A – a benchmark based on multiple-choice questions in various scientific fields.

According to OpenAI, the new and improved ChatGPT is “more direct” and “less verbose” too, and will use “more conversational language”. All in all, a bit more human-like then. Eventually, the improvements should trickle down to non-paying users too.

More up to date

See more

In an example given by OpenAI, AI-generated text for an SMS intended to RSVP to a dinner invite is half the length and much more to the point – with some of the less essential words and sentences chopped out for simplicity.

Another important upgrade is that the training data ChatGPT is based on now goes all the way up to December 2023, rather than April 2023 as with the previous model, which should help with topical questions and answers.

It's difficult to test AI chatbots from version to version, but in our own experiments  with ChatGPT and GPT-4 Turbo we found it does now know about more recent events – like the iPhone 15 launch. As ChatGPT has never held or used an iPhone though, it's nowhere near being able to offer the information you'd get from our iPhone 15 review.

The momentum behind AI shows no signs of slowing down just yet: in the last week alone Meta has promised human-like cognition from upcoming models, while Google has made its impressive AI photo-editing tools available to more users.

You might also like

TechRadar – All the latest technology news

Read More

ChatGPT’s new AI store is struggling to keep a lid on all the AI girlfriends

On Jan 10, OpenAI officially launched its GPT Store, thus opening the door for select users and official partners to create, search for, and try out customized ChatGPTs (read: Chat Generative Pre-trained Transformers – or AI chatbots) provided they subscribe to either ChatGPT Plus for $ 20 / £16 / AU$ 20 a month, Enterprise, or the new Teams plan. 

And despite OpenAI's usage policies explicitly stating “We also don’t allow GPTs dedicated to fostering romantic companionship or performing regulated activities”, alongside warnings to creators such as: “Don't build tools that may be inappropriate for minors, including sexually explicit or suggestive content”, after just seven days the Store has seemingly been flooded with virtual girlfriends. 

Both Quartz and subsequently Mashable published images of searches for “girlfriend” and later “sweetheart” (by Mashable) in the GPT Store yielding many results, although some hits had disappeared in the four days between reports. 

For us, the search bar within the GPT Store (which we might use to search for “girlfriend”, “sweetheart,” “flirting” and so on) has seemingly been removed, thus suggesting OpenAI is trying to get a handle on the situation – although we do still see a tab to 'Create a GPT' in Beta. 

Analysis: users want AI girlfriends, so creators will do what they do best – get creative 

A screen-grab of the GPT Store's landing page, with no search bar

Note the lack of search bar for us in the GPT Store (Image credit: GPT Store)

So how easy is it to get an AI girlfriend on OpenAI's GPT Store, despite these kinds of chatbots directly contradicting the company's clear Ts&Cs? 

Although the store’s page isn't currently allowing us to search for virtual romantic partners (or any kind of GPT) because the search bar just isn't there right now, third-party sites will still let you perform these kinds of searches, with some links going to the actual GPT Store. 

I tried “girlfriend” (plenty of options) and even “secret lover” on the site GPTStore.ai, which yielded a GPT described as a “Romantic AI partner for text adventure dates with image creation”, albeit linking to the author's website rather than to the official GPT Store. And yes, I am now waiting for the call from IT about my questionable searches today… 

The GPT Store was originally announced last November as part of the company’s first DevDay conference, alongside OpenAI’s then-new create-a-chatbot service. The store was slated to open by the end of that month but was delayed several times, likely in part due to the somewhat abrupt dismissal and reinstatement of CEO Sam Altman.  

There is of course a wider discussion to be had here regarding the very real danger of users developing romantic attachments to AI. Business Insider reported on Monday (January 15) that the recent surge in platforms dedicated to AI companions has seen chatbot app Replika (described as an “AI for anyone who wants a friend with no judgment, drama, or social anxiety involved”) downloaded more than 10 million times.

You might also like

TechRadar – All the latest technology news

Read More

Microsoft’s Copilot chatbot will get 6 big upgrades soon – including ChatGPT’s new brain

Microsoft has announced that Copilot – the AI chatbot formerly known as Bing Chat – is soon to get half a dozen impressive upgrades.

This batch of improvements should make the Copilot chatbot considerably more powerful in numerous respects (outside of, and inside of Windows 11).

So, first off, let’s break down the upgrades themselves (listed in a Microsoft blog post) before getting into a discussion of what difference they’re likely to make.

Firstly, and most importantly, Copilot is getting a new brain, or we should say an upgraded brain in the form of GPT-4 Turbo. That’s the latest model of GPT from OpenAI which makes various advances in terms of being generally better and more accurate.

Another beefy upgrade is an updated engine for Dall-E 3, the chatbot’s image creation feature, which produces higher quality results that are more closely aligned with what’s requested by the user. This is actually in Copilot right now.

Thirdly, Microsoft promises that Copilot will do better with image searches, returning better results when you sling a picture at the AI in order to learn more about it.

Another addition is Deep Search which uses GPT-4 to “deliver optimized search results for complex topics” as Microsoft puts it. What this means is that if you have a query for Copilot, it can produce a more in-depth search request to produce better results. Furthermore, if the terms of your query are vague and could potentially relate to multiple topics, Deep Search will follow up on what those topics might be and offer suggestions to allow you to refine the query.

The fifth upgrade Microsoft has planned is Code Interpreter, which as the name suggests will help perform complex tasks including coding, data analysis and so forth. That’s not something the average user will benefit from, but there are those who will, of course.

Finally, Copilot in Microsoft’s Edge browser has a rewrite feature (for inline text composition) coming soon. This allows you to select text on a website and get the AI to rewrite it for you.


Analysis: Something for Google to worry about

Dall-E 3

(Image credit: Future)

There are some really useful changes inbound here. Getting GPT-4 Turbo is an upgrade (from GPT-4) that a lot of Copilot users have been clamoring for, and Microsoft tells us that it’s now being tested with select users. (We previously heard it still had a few kinks to iron out, so presumably that’s what’s currently going on).

GPT-4 Turbo will be rolling out in the “coming weeks” so it should be here soon enough, with any luck, and you’ll be able to see the difference it makes for you in terms of a greater level of accuracy for the chatbot when responding to your queries.

It’s great to see Dall-E 3 getting an upgrade, too, as it’s already an excellent image creation engine, frankly. (Recall the rush to use the feature when it was first launched, due to the impressive results being shared online).

The search query improvements, both the Deep Search capabilities and refined image searching, will also combine with the above upgrades to make Copilot a whole lot better across multiple fronts. (Although we do worry somewhat about the potential for abuse with that inline rewrite feature for Edge).

All of this forward momentum for Copilot comes as we just heard news of Google delaying its advances on the AI front, pushing some major launches back to the start of 2024. Microsoft isn’t hanging around when it comes to Copilot, that’s for sure, and Google has to balance keeping up, without pushing so hard that mistakes are made.

You might also like …

TechRadar – All the latest technology news

Read More

Google denies that Bard AI copied ChatGPT’s homework

Google’s Bard AI has found itself at the center of controversy again, this time over allegations that the Bing rival was trained using data pulled from OpenAI’s ChatGPT.

As you may be aware, ChatGPT is the power behind the throne of Bing AI, and the accusation of nefarious activities behind the scenes comes from a report by The Information.

We’re told that Jacob Devlin, a software engineer at Google – an ex-engineer, we might add, having departed the firm over this affair – claims that Google used ChatGPT data (scraped from the ShareGPT website, apparently) to develop Bard.

Devlin notes that he warned Google against doing so, as this clearly went against OpenAI’s terms of service.

According to the report, Google ceased using the mentioned data after the warnings from Devlin (who left Google to join OpenAI, we’re informed).

Google denies any of this, though. A company spokesperson, Chris Pappas, told The Verge: “Bard is not trained on any data from ShareGPT or ChatGPT.”


Analysis: A denial amid some desperation

There we have it, then – a clear denial from Google in no uncertain terms that nothing underhand was going on data-wise with Bard. And to be fair, there’s certainly no evidence that Bard’s answers are remotely like the ones given by ChatGPT. (Devlin had further warned that the alleged data hoovering could mean just that, and it’d be obvious enough what had gone on as a result).

We suppose the trouble with this episode is that it very much feels like Google has rushed Bard to release – dropping clangers while doing so – as it was forced to play catchup with Microsoft’s Bing AI. Given that the latter is now successfully pushing search engine adoption to Bing, already at this early stage, all this could make it easy enough for some to believe that Google might be getting a bit desperate with tactics behind the scenes.

Whether or not the tale about poached data is true – we’ll take Google’s word that it isn’t – the report still makes an interesting revelation that Google’s Brain AI group is now working with AI firm DeepMind (both of these existing under the Alphabet umbrella, the parent company).

DeepMind has seemingly been recruited into the mix to swiftly hone and power up Bard, and it’s notable because the two AI outfits are big rivals and are very much being forced to collaborate on this.

This again sketches a picture of a rather desperate scramble to get Bard steadier on its feet, while Microsoft’s Bing AI keeps getting updated with new features at a fair old rate of knots. (Although fresh rumblings about one of the potential next ‘features’ for the Bing chatbot have us very concerned, it has to be said).

You may also recall alarm bells being rung on the privacy front when Bard itself made an apparent revelation that it used internal Gmail data for training, again prompting Google to tell us that this is not the case and that the bot got things wrong. Bard getting things wrong, of course, is very much part of a bigger issue.

TechRadar – All the latest technology news

Read More