Don't be fooled: Google faked its Gemini AI voice demo

Trending 2 months ago

AI In brief Google wowed the internet with a audience video assuming the multimodal capabilities of its latest ample accent archetypal Gemini – but some of the the audience was faked.

In the audience below, Gemini seems to be able to acknowledge to a user's articulation and collaborate with a user's surroundings, attractive at things they accept fatigued or arena rock, paper, scissors. In the demo, Gemini is asked to assumption what the user is abstraction on a Post-It agenda and accurately answers duck, for example.

A elastic avoid is again placed on a cardboard album and Gemini is able to analyze area the article has been placed. It does all sorts of things – anecdotic objects, award area things accept been hidden and switched beneath cups, and more. Google approved to appearance off Gemini's abilities to action altered forms of information, and accomplish analytic and spatial reasoning. 

Youtube Video

But in reality, the archetypal was not prompted application audio and its responses were alone text-based. They were not generated in absolute time either. Instead, the video was crafted "using still angel frames from the footage, and bidding via text," a Google agent told Bloomberg. 

The being speaking in the audience was absolutely account out some of the argument prompts that were anesthetized to the model, and the apprentice articulation accustomed to Gemini was account out responses it had generated in text. Still images taken from the video – like the rock, paper, scissors – were fed to the model, and it was asked to assumption the game. Google again cherry-picked its best outputs and anecdotal them alongside the footage to accomplish it assume as if the archetypal could acknowledge flawlessly in absolute time.

"For the purposes of this demo, cessation has been bargain and Gemini outputs accept been beneath for brevity," the description for the video on YouTube reads. Oriol Vinyals, VP of analysis and abysmal acquirements advance at Google DeepMind, who helped advance the Gemini project, admitted that the video demonstrates "what the multimodal user adventures complete with Gemini could attending like" (our emphasis).

AMD is advancing for Nvidia's lunch

Top AI developers accept committed to application AMD's latest Instinct MI300-series accelerators as they attending for added computational assets to abutment the training and active of their models. 

At AMD's barrage this week, admiral from Microsoft, Oracle, and Supermicro went on date to abutment the cavity shop, alliance to acquirement and body AI servers to ability billow platforms, or standalone machines. Microsoft will use the chips to body MI300x v5 Virtual Machine clusters for Azure, while Oracle will action OCI bald metal compute solutions.

Dell will accommodate AMD's latest AI accelerators for its PowerEdge XE9680 servers, while HPE will alpha to arrange them for its HPC business. Meanwhile, added Meta promised to add the chips to its datacenters, and OpenAI is developing software to abutment the Instinct MI300 application its Triton 3.0 compiler. 

"AI is the approaching of accretion and AMD is abnormally positioned to ability the end-to-end basement that will ascertain this AI era, from massive billow installations to action clusters and AI-enabled able anchored accessories and PCs," AMD CEO Lisa Su declared in a statement.

Nvidia is at the beginning of AI compute, and its revenues accept developed massively year over year as appeal for its GPUs rises. But accumulation is short, and big barter are attractive for added options. Some with the centermost pockets accept alike angry to architecture their own custom silicon – like Google, Amazon, and Microsoft. 

It's a acceptable time to try and abduct some of Nvidia's lunch, and AMD's Instinct MI300 alternation is its best attack so far. As added and added developers accept the chip, the software ecosystem advised to abutment its accouterments will abound – authoritative it easier for others to use AMD's hardware.

  • Google launches Gemini AI systems, claims it's assault OpenAI and others - mostly
  • AMD slaps calm a silicon sandwich with MI300-series APUs, GPUs to claiming Nvidia's AI empire
  • Strike over? US actors may acknowledgment to assignment with top-tier 'progressive AI protections'
  • Meta trials Purple Llama activity for AI developers to analysis assurance risks in models

SAG-AFTRA associates vote to accept abutment arrangement acclimation AI

US abecedarian abutment SAG-AFTRA has clearly ratified its acceding with top TV and blur assembly companies afterwards extensive a accord over bigger alive altitude and AI.

Members concluded their months-long bang and alternate to alive back leaders managed to accommodate bigger arrangement agreement with the Alliance of Motion Picture and Television Producers (AMPTP). A big afraid point was acclimation the use of AI as the technology becomes added beat and is adopted added broadly by the ball industry.

Under the deal, media studios charge access absolute accord and atone performers for application their likeness. Actors and actresses were anxious that they could be replaced and lose out on jobs to companies axis to technology to actualize apocryphal but realistic-looking account or choir for adverts, TV shows, or films.  

The acceding was formally ratified afterwards the majority of associates voted in favor of it this week.

"SAG-AFTRA associates accepted a axiological change in the way this industry treats them: candor in advantage for their labor, aegis from calumniating use of AI technology, adequate account plans, and candid and admiring analysis for all members, amid added things," the union's civic controlling administrator & arch adjudicator Duncan Crabtree-Ireland explained in a statement. 

"This new arrangement delivers on these objectives and makes abundant advance in affective the industry in the appropriate direction. By acknowledging this contract, associates accept fabricated it bright that they're acquisitive to use their accord to lay the background for a bigger industry, convalescent the lives of those alive in their profession."

Meta releases text-to-image apparatus and promises to watermark its images

Meta appear Imagine – a web-based text-to-image app – this week, and is planning to add a agenda watermark to characterization constructed agreeable generated by its software.

Imagine is powered by Emu, which is a beheld abundant AI archetypal able of creating 2D- and abbreviate 3D-animated videos. It can be acclimated by anyone with a Facebook account. Type in a abbreviate prompt, and Imagine will accomplish a console of still images analogous the ascribe description that users can flick through and use.

Meta is planning to cycle out technology that automatically adds a watermark to Imagine's outputs to accomplish abiding the AI-generated agreeable can be detected.

"In the advancing weeks, we'll add airy watermarking to [Imagine] with Meta AI acquaintance for added accuracy and traceability. The airy watermark is activated with a abysmal acquirements model. While it's ephemeral to the animal eye, the airy watermark can be detected with a agnate model," Meta accepted in a blog post.

The amusing belvedere claimed that the watermark will abide complete alike if users crop, alter, or booty screenshots of Imagine's AI-generated images. ®