Rabbit r1 | Have we finally created a gadget that can eat your smartphone?

This year’s Consumer Electronics Show at Las Vegas was littered with updates from both start-ups and large tech firms that are building products harnessing, or in some cases, advancing the power of natural language processing (NLP), a burgeoning sub-field under artificial intelligence (AI).

With so many exhibits, it is difficult to point out any one piece of tech as exceptional this year. Still, an orange-coloured, square-shaped device unveiled at the ballroom at Wynn, and not at the official CES stage, grabbed the spotlight.

The palm-sized handheld, called Rabbit r1, received a fair amount of chatter at CES 2024 as it could do – – per the company’s claim – – several things that a smartphone can’t. Even Microsoft CEO Satya Nadella called it the ‘most impressive’ device, and compared it to the first iPhone unveiled by Steve Jobs.

So, what exactly does this device do?

If you want to book an Uber ride, the r1 can do it for you. If you want to plan a vacation, including booking air tickets and making room reservations, the r1 can do that for you. If you want some cooking ideas, the r1’s camera can scan the motley ingredients in your refrigerator and suggest a recipe based on your calorie requirement. All you have to do is just ‘tell it’ what to do.

(For top technology news of the day, subscribe to our tech newsletter Today’s Cache)

Exploiting chatbots’ limitation

Granted, any of the latest generation smartphones with its state-of-the-art voice assistant can do several tasks like searching the web, playing your favourite song, or making a call from a user’s phonebook. But, executing tasks, like booking a cab, reserving hotel room, and putting together a recipe using computer vision, just by talking into a walkie-talkie style device, is a stretch even for smartphone-based voice assistants.

Even the current crop of chatbots, like ChatGPT, Bard and Claude, can only text out responses through apps as they are incapable of executing actionable tasks. For instance, the ChatGPT app can text you a vacation plan. It can even tweak the itinerary if you ask it to make it easy or packed. But, it cannot open a ticket booking app or a room reservation portal to make a reservation for you.

Rabbit Inc., the maker of r1, says that the current batch of chatbots have limited functionality because they are built on text-based AI models – – more commonly known as large language models (LLMs). LLMs’ accuracy depends a lot on annotated data to train neural networks for every new task.

Extending LLM’s capability

The Santa Monica-based start-up, on the other hand, has built its r1 device using a different AI model that is biased for action. The Rabbit OS, in a way, extends the capabilities of the current generation of voice assistants.

The AI model, which the company calls a large action model (LAM), takes advantage of advances in neuro-symbolic programming, a method that combines the data driven capabilities of the neural networks with symbolic reasoning techniques. This allows the device to directly learn from the user’s interaction with the applications and execute tasks, essentially bypassing the need to translate text-based user requests into APIs.

Apart from bypassing the API route, LAM-based OS caters to a more nuanced human to machine interaction. While ChatGPT can be creative in responding to prompts, a LAM-based OS learns routine and minimalistic tasks with a sole purpose of repeating it.

So, Rabbit Inc., in essence, has created a platform, underpinned by an AI model, that can mimic what humans do with their smartphones and then repeat it when asked to execute. The r1 is the company’s first generation device, which according to its founder Jesse Lyu, is a stand-alone gadget that is primarily driven by natural language “to get things done.”

The company has also cleverly priced the device at $199, significantly lesser than the price of most flagship smartphones. This makes it difficult to decipher whether customers will buy this device for the value it offers or just because it is cheap.

But is the price differentiation alone enough to trade in your existing smartphone for the new Rabbit r1?

A smartphone replacement?

Booking a ride, planning a vacation, or playing music are only a subset of things we do with a smartphone. Over last roughly one and half decade the smartphone has become a pocket computer.

The app ecosystem built for this hardware has made the device so sticky that an average user picks up their smartphone at least 58 times a day, and spends, on average, at least three hours with it. And during that time, they use this mini-computer for whole host of things, not to mention streaming videos, playing games, reading books, and interacting with friends and family via group chat applications.

Secondly, not everyone wants to speak into a device all the time to get something done. Most people are just fine typing in text prompts and getting responses in the same format. It gives them a layer of privacy that the r1 does not provide – – that’s because the latter can only execute voice commands.

So, the smartphone, and its app ecosystem, is here to stay to cater to an entire gamut of user needs and wants for the foreseeable future.

Now, where does that leave Rabbit r1?

Into the Rabbit hole

Mr. Lyu believes the r1 will disrupt the smartphone market, but technically, his company’s palm-sized device is a strong contender in the voice assistant and smart speaker market, which is also space that is growing quite steadily.

According to a 2022 joint report by NPR and Edison Research, in the U.S. alone, 62% of users over the age of 18 use voice assistant on any smart device. And the number of tasks they do with it is alap increasing: In 2022, smart speaker users requested an average of 12.4 tasks on their device each week, up from 7.5 in 2017. And smartphone voice assistant users requested an average of 10.7 tasks weekly, up from 8.8 in 2020.

This shows that the r1 can play an important transition role in the audio space by driving hardware designers and software developers in the direction of building more voice-based interoperable application. Alternatively, Rabbit inc can also building a super app — something like a WeChat app that can enable chatter between apps in a smarphone to ‘get things done.’

That’s a call Rabbit Inc. should take based on the feedback it receives from its customers. As on January 19, five batches of 10,000 (batch size) rabbit r1 devices have been sold out. And the first batch will start shipping in April. Customer experience with this new gadget will play a big role in how deep r1 will take consumers down the rabbit hole.

Source link

#Rabbit #finally #created #gadget #eat #smartphone

Nvidia and AI changed landscape of the chip industry, as rivals play catch-up

This year’s artificial-intelligence boom turned the landscape of the semiconductor industry on its head, elevating Nvidia Corp. as the new king of U.S. chip companies — and putting more pressure on the newly crowned company for the year ahead.

Intel Corp.
which had long been the No. 1 chip maker in the U.S., first lost its global crown as biggest chip manufacturer to TSMC

several years ago. Now, Wall Street analysts estimate that Nvidia’s

annual revenue for its current calendar year will outpace Intel’s for the first time, making it No. 1 in the U.S. Intel is projected to see 2023 revenue of $53.9 billion, while Nvidia’s projected revenue for calendar 2023 is $56.2 billion, according to FactSet.

Even more spectacular are the projections for Nvidia’s calendar 2024: Analysts forecast revenue of $89.2 billion, a surge of 59% from 2023, and about three times higher than 2022. In contrast, Intel’s 2024 revenue is forecast to grow 13.3% to $61.1 billion. (Nvidia’s fiscal year ends at the end of January. FactSet’s data includes pro-forma estimates for calendar years.)

“It has coalesced into primarily an Nvidia-controlled market,” said Karl Freund, principal analyst at Cambrian AI Research. “Because Nvidia is capturing market share that didn’t even exist two years ago, before ChatGPT and large language models….They doubled their share of the data-center market. In 40 years, I have never seen such a dynamic in the marketplace.”

Nvidia has become the king of a sector that is adjacent to the core-processor arena dominated by Intel. Nvidia’s graphics chips, used to accelerate AI applications, reignited the data-center market with a new dynamic for Wall Street to watch.

Intel has long dominated the overall server market with its Xeon central processor unit (CPU) family, which are the heart of computer servers, just as CPUs are also the brain chips of personal computers. Five years ago, Advanced Micro Devices Inc.
Intel’s rival in PC chips, re-entered the lucrative server market after a multi-year absence, and AMD has since carved out a 23% share of the server market, according to Mercury Research, though Intel still dominates with a 76.7% share.

Graphics chips in the data center

Nowadays, however, the data-center story is all about graphics processing units (GPUs), and Nvidia’s have become favored for AI applications. GPU sales are growing at a far faster pace than the core server CPU chips.

Also read: Nvidia’s stock dubbed top pick for 2024 after monster 2023, ‘no need to overthink this.’

Nvidia was basically the entire data-center market in the third quarter, selling about $11.1 billion in chips, accompanying cards and other related hardware, according to Mercury Research, which has tracked the GPU market since 2019. The company had a stunning 99.7% share of GPU systems in the data center, excluding any devices for networking, according to Dean McCarron, Mercury’s president. The remaining 0.3% was split between Intel and AMD.

Put another way: “It’s Nvidia and everyone else,” said Stacy Rasgon, a Bernstein Research analyst.

Intel is fighting back now, seeking to reinvigorate growth in data centers and PCs, which have both been in decline after a huge boom in spending on information technology and PCs during the pandemic. This month, Intel unveiled new families of chips for both servers and PCs, designed to accelerate AI locally on the devices themselves, which could also take some of the AI compute load out of the data center.

“We are driving it into every aspect of the applications, but also every device, in the data center, the cloud, the edge of the PC as well,” Intel CEO Pat Gelsinger said at the company’s New York event earlier this month.

While AI and high-performance chips are coming together to create the next generation of computing, Gelsinger said it’s also important to consider the power consumption of these technologies. “When we think about this, we also have to do it in a sustainable way. Are we going to dedicate a third, a half of all the Earth’s energy to these computing technologies? No, they must be sustainable.”

Meanwhile, AMD is directly going after both the hot GPU market and the PC market. It, too, had a big product launch this month, unveiling a new family of GPUs that were well-received on Wall Street, along with new processors for the data center and PCs. It forecast it will sell at least $2 billion in AI GPUs in their first year on the market, in a big challenge to Nvidia.

Also see: AMD’s new products represent first real threat to Nvidia’s AI dominance.

That forecast “is fine for AMD,” according to Rasgon, but it would amount to “a rounding error for Nvidia.”

“If Nvidia does $50 billion, it will be disappointing,” he added.

But AMD CEO Lisa Su might have taken a conservative approach with her forecast for the new MI300X chip family, according to Daniel Newman, principal analyst and founding partner at Futurum Research.

“That is probably a fraction of what she has seen out there,” he said. “She is starting to see a robust market for GPUs that are not Nvidia…We need competition, we need supply.” He noted that it is early days and the window is still open for new developments in building AI ecosystems.

Cambrian’s Freund noted that it took AMD about four to five years to gain 20% of the data-center CPU market, making Nvidia’s stunning growth in GPUs for the data center even more remarkable.

“AI, and in particularly data-center GPU-based AI, has resulted in the largest and most rapid changes in the history of the GPU market,” said McCarron of Mercury, in an email. “[AI] is clearly impacting conventional server CPUs as well, though the long-term impacts on CPUs still remain to be seen, given how new the recent increase in AI activity is.”

The ARMs race

Another development that will further shape the computing hardware landscape is the rise of a competitive architecture to x86, known as reduced instruction set computing (RISC). In the past, RISC has mostly made inroads in the computing landscape in mobile phones, tablets and embedded systems dedicated to a single task, through the chip designs of ARM Holdings Plc

and Qualcomm Inc.

Nvidia tried to buy ARM for $40 billion last year, but the deal did not win regulatory approval. Instead, ARM went public earlier this year, and it has been promoting its architecture as a low-power-consuming option for AI applications. Nvidia has worked for years with ARM. Its ARM-based CPU called Grace, which is paired with its Hopper GPU in the “Grace-Hopper” AI accelerator, is used in high-performance servers and supercomputers. But these chips are still often paired with x86 CPUs from Intel or AMD in systems, noted Kevin Krewell, an analyst at Tirias Research.

“The ARM architecture has power-efficiency advantages over x86 due to a more modern instruction set, simpler CPU core designs and less legacy overhead,” Krewell said in an email. “The x86 processors can close the gap between ARM in power and core counts. That said, there’s no limit to running applications on the ARM architecture other than x86 legacy software.”

Until recently, ARM RISC-based systems have only had a fractional share of the server market. But now an open-source version of RISC, albeit about 10 years old, called RISC-V, is capturing the attention of both big internet and social-media companies, as well as startups. Power consumption has become a major issue in data centers, and AI accelerators use incredible amounts of energy, so companies are looking for alternatives to save on power usage.

Estimates for ARM’s share of the data center vary slightly, ranging from about 8%, according to Mercury Research, to about 10% according to IDC. ARM’s growing presence “is not necessarily trivial anymore,” Rasgon said.

“ARM CPUs are gaining share rapidly, but most of these are in-house CPUs (e.g. Amazon’s Graviton) rather than products sold on the open market,” McCarron said. Amazon’s

 Graviton processor family, first offered in 2018, is optimized to run cloud workloads at Amazon’s Web Services business. Alphabet Inc.


also is developing its own custom ARM-based CPUs, codenamed Maple and Cypress, for use in its Google Cloud business according to a report earlier this year by the Information.

“Google has an ARM CPU, Microsoft has an ARM CPU, everyone has an ARM CPU,” said Freund. “In three years, I think everyone will also have a RISC-V CPU….It it is much more flexible than an ARM.”

In addition, some AI chip and system startups are designing around RISC-V, such as Tenstorrent Inc., a startup co-founded by well-regarded chip designer Jim Keller, who has also worked at AMD, Apple Inc.
Tesla Inc.

and Intel.

See: These chip startups hope to challenge Nvidia but it will take some time.

Opportunity for the AI PC

Like Intel, Qualcomm has also launched an entire product line around the personal computer, a brand-new endeavor for the company best known for its mobile processors. It cited the opportunity and need to bring AI processing to local devices, or the so-called edge.

In October, it said it is entering the PC business, dominated by Intel’s x86 architecture, with its own version of the ARM architecture called Snapdragon X Elite platform. It has designed its new processors specifically for the PC market, where it said its lower power consumption and far faster processing are going to be a huge hit with business users and consumers, especially those doing AI applications.

“We have had a legacy of coming in from a point where power is super important,” said Kedar Kondap, Qualcomm’s senior vice president and general manager of compute and gaming, in a recent interview. “We feel like we can leverage that legacy and bring it into PCs. PCs haven’t seen innovation for a while.”

Software could be an issue, but Qualcomm has also partnered with Microsoft for emulation software, and it trotted out many PC vendors, with plans for its PCs to be ready to tackle computing and AI challenges in the second half of 2024.

“When you run stuff on a device, it is secure, faster, cheaper, because every search today is faster. Where the future of AI is headed, it will be on the device,” Kondap said. Indeed, at its chip launch earlier in this month, Intel quoted Boston Consulting Group, which forecast that by 2028, AI-capable PCs will comprise 80% of the PC market..

All these different changes in products will bring new challenges to leaders like Nvidia and Intel in their respective arenas. Investors are also slightly nervous about Nvidia’s ability to keep up its current growth pace, but last quarter Nvidia talked about new and expanding markets, including countries and governments with complex regulatory requirements.

“It’s a fun market,” Freund said.

And investors should be prepared for more technology shifts in the year ahead, with more competition and new entrants poised to take some share — even if it starts out small — away from the leaders.

Source link

#Nvidia #changed #landscape #chip #industry #rivals #play #catchup

Lukas Gage’s viral video audition haunts the ‘hot labor summer’ actors’ strike sweeping Hollywood

In November 2020, the actor Lukas Gage was auditioning for a role via video link when he heard the producer make some disparaging remarks about the size of his apartment. 

“These poor people who live in these tiny apartments,” the producer said. “I’m looking at his background and he’s got his TV and …”

Gage, who at that time had had a four-episode arc on HBO’s “Euphoria” among other small roles, interrupted the producer — British director Tristram Shapeero, who later apologized for his remarks — to let him know that he was not muted and that Gage could, in fact, hear him. 

“Yeah, I know it’s a sh—y apartment,” Gage said. “That’s why — give me this job so I can get a better one.”

Shapeero replied, “Oh my god, I am so, so sorry … I am absolutely mortified.”

Putting together an audition tape can often take up an entire day and involve setting up a studio space for sound and lighting.

“Listen, I’m living in a four-by-four box, just give me the job and we’ll be fine,” Gage responded. 

Gage kept his sense of humor, but he also decided to post the video on his Twitter account to show how actors are sometimes treated from the moment they audition for a role — and perhaps to remind people to make sure you’re on mute if you’re trash-talking someone on a Zoom


It’s three years later, and members of the Writers Guild and Screen Actors Guild are on strike, looking for more pay, better working conditions and stricter rules around things like the use of actors’ images in the age of artificial intelligence and the lack of residuals from streaming networks. 

The perils of the online audition

Meanwhile, Gage’s 2020 online audition is resonating again. 

For a working actor — who, like the majority of SAG-AFTRA members who may not be an A-list star — simply getting in front of a producer as Gage did can be a long and difficult process. And since the start of the pandemic, the nature of auditions has changed dramatically. This has come to symbolize the uphill struggle actors face from the moment they hear about a role. 

In May, Ezra Knight, New York local president of SAG-AFTRA, asked members to authorize strike action, saying contracts needed to be renegotiated to reflect dramatic changes in the industry. Knight cited the need to address artificial intelligence, pay, benefits, reduced residuals in streaming and “unregulated and burdensome self-taped auditions.”

In the days of live auditions, actors would read for a role with a casting director. But several actors told MarketWatch that it’s become harder to make a living in recent years, and that it all starts with the audition tape, which has now become standard in the industry. 

By the time Gage got in front of producers, for instance, he had likely either already delivered a tape and was put on a shortlist to read in front of a producer, or the casting director was already familiar with his work and wanted him to read for the part. 

But an audition tape can often take up an entire day to put together, actors say. When the opportunity to audition arrives, actors typically have to drop everything they’re doing — whether they’re working a side hustle or taking time off or even enjoying a vacation.

Cadden Jones: “All the financial responsibilities have fallen on us. The onus is on us to create our auditions.”

Cadden Jones

They need to arrange good lighting and a clean backdrop — Gage’s TV set became a distraction for the producer during his audition — set up the camera, and scramble to find a “reader” — someone to read the other roles in the scene, preferably another actor. 

Then the actor has to edit the audition to highlight their strongest take and upload it. There are currently no regulations on the amount of pages a casting director can send to a candidate, and actors say there’s often not enough time to properly prepare.

“Unfortunately, it’s been going in this direction for some time now,” said Cadden Jones, an actor based in New York who has credits on shows including Showtime’s

“Billions” and Amazon Prime’s

“The Marvelous Mrs. Maisel.” 

“This was the first year I did not qualify for health insurance in decades,” she told MarketWatch. “I just started teaching.”

To put that into perspective: Members of SAG-AFTRA must earn $26,470 in a 12-month base period to qualify for health insurance. The median annual wage in the U.S. hovers at around $57,000, based on the weekly median as calculated by the Bureau of Labor Statistics.

Jones and her partner, Michael Schantz, an actor who works mostly in theater, are starting a communications consulting company to increase their income.

“Most if not all of my actor friends have had to supplement their income since the pandemic,” she said. “We’re in trouble as a community of actors who used to make a good living doing what we do. It’s not like any of us lost our talent overnight. I, for one, am very glad that we’re striking.”

But Jones said that, with the auditioning process taking place mostly online since the onset of the pandemic, casting agents — who work for producers — are able to see more people for a given role, making the competition for roles even more intense.

‘This was the first year I did not qualify for health insurance in decades.’

— Cadden Jones, an actor based in New York

“We don’t go into casting offices anymore,” Jones said. “All the financial responsibilities have fallen on us. The onus is on us to create our auditions. It’s harder to know what they want, and you don’t have the luxury to work with a casting director in a physical space to get adjustments, which was personally my favorite part of the process — that collaboration.”

She added: “Because the audition rate accelerated, the booking rate went down dramatically for everybody. But don’t get me wrong. Once the strike is officially over, I want all the auditions I can get.”

SAG-AFTRA has proposed rules and expectations to address some of the burden and costs actors bear when it comes to casting, including providing a minimum amount of time for actors to send in self-taped auditions; disclosing whether an offer has been made for the role or it has already been cast; and limiting the number of pages for a “first call” or first round of auditions.

Before the negotiations broke down with the actors’ union, the Alliance of Motion Picture and Television Producers, which represents over 350 television and production companies, said it offered SAG-AFTRA $1 billion in wage increases, pension and health contributions and residual increases as part of a range of proposals related to pay and working conditions.

Those proposals included limitations on requests for audition tapes, including page, time and technology requirements, as well as options for virtual or in-person auditions, AMPTP said. The producers’ group characterized their offer as “the most lucrative deal we have ever negotiated.”

Michael Schantz: “How does the broader culture value storytelling and the people who make stories?”

Michael Schantz

Jones said she doesn’t blame the casting directors. It’s up to the producers, she said, to be more mindful of how the changes in the industry since the advent of streaming, the decline in wages adjusted for inflation, and poor residuals from streaming services have taken a toll on working actors.

Bruce Faulk, who has been a member of SAG-AFTRA since 1992, said that for work on a one-off character part or a recurring role on a network show, he might receive a check for hundreds or even thousands of dollars in residuals. And — crucially — he knows how many times a particular show has aired. 

Residuals are fees paid to actors each time a TV show or film is broadcast on cable or network television. They are based on the size of the role and the budget of the production, among other things. For shows that air on streaming services, however, residuals are far harder to track. 

What’s more, residuals decline over time and can often amount to just a few cents per broadcast. 

Actor Kimiko Glenn, who appeared on episodes of Netflix’s

“Orange Is the New Black,” recently shared a video on TikTok showing $27 in residuals from her work on that show.

Faulk sympathizes. “A lot of checks from HBO

for ‘The Sopranos’ or ‘Gossip Girl’ I get are for $33,” he said. “I never know how many people watched me on ‘Gossip Girl’ in the three episodes I’m in. All we know is whatever the streaming services decided to announce as their subscriber numbers.”

Like Jones, Faulk said this will be the first year he won’t qualify for SAG-AFTRA health insurance, which covers him, his wife and his son. This is despite him having worked enough over the past 10 years to qualify for a pension when he turns 67. “Mine is up to $1,000 a month now,” he said, noting that the pension will keep increasing if he keeps getting acting work.

Schantz, who had a three-episode arc on NBC’s

“The Blacklist” in addition to his other TV, film and theater credits, finds the recent shifts in the landscape for actors somewhat difficult to reconcile with the way people turned to TV and film during the loneliest days of the pandemic.

“One of the most concerning things I can think of right now is the conversation around value. How does the broader culture value storytelling and the people who make stories?” he said. “The arts always tend to fall to the wayside in many ways, but it was striking during the pandemic that so much of our attention went to watching movies and television. There’s obviously something inside of us that feels like we’re part of the human story.”

Actors battle other technology

While big companies like Disney
HBO, Apple
Amazon and Netflix make millions of dollars from films and TV series that are watched again and again, Schantz said that actors are unable to make a living. “No one wants to go on strike,” he said. 

Those five companies have not responded to requests for comment from MarketWatch on these issues.

Since his audition tape went viral, Gage has booked regular work, and he found even greater fame when he went on to star in Season 1 of HBO’s “White Lotus.” In 2023, he will star in nine episodes of “You,” now streaming on Netflix, and in the latest season of FX’s “Fargo.” 

Earlier this year, he told the New York Times: “I had never judged my apartment until that day.” He added, “I remember having this weird feeling in the pit of my stomach afterward, like, why am I judging where I’m at in my 20s, at the beginning of my career?”

‘There’s enough Bruce out there where you could take my likeness and my voice and put me in the scene.’

— Bruce Falk, a member of SAG-AFTRA since 1992

But advances in technology are not just hurting actors in the audition process. A debate is raging over the use of AI and whether actors should be expected to sign away the rights to their image in perpetuity, especially when they might only be getting paid for half a day’s work.

“AI is the next big thing,” Falk said. The industry is concerned about companies taking actors’ likenesses and using AI to generate crowd scenes. 

“Even an actor at my level — that guy on that show — there’s enough Bruce out there where you could take my likeness and my voice and put me in the scene: the lieutenant who gives you the overview of what happened to the dead body,” he said. “At this point, I could be technically replaced. We have to get down on paper, in very clear terms, that that can’t be done.”

The Alliance of Motion Picture and Television Producers also said it agrees with SAG-AFTRA and had proposed — before the actors’ strike — “that use of a performer’s likeness to generate a new performance requires consent and compensation.” The AMPTP said that would mean no digital version of a performer should be created without the performer’s written consent and a description of the intended use in the film, and that later digital replicas without that performer’s consent would be prohibited.  

“Companies that are publicly traded obviously have a fiduciary responsibility to their shareholders, and whatever they can use, they will use it — and they are using AI,” Schantz said. “Yes, there are some immediate concerns. Whether or not the technology is advanced enough to fully replace actors is an open question, but some people think it’s an inevitability now.

“To let companies have free rein with these technologies is obviously creating a problem,” he added. “I can’t go show up, do a day’s work, have my performance be captured, and have that content create revenue for a company unless I’m being property compensated for it.”

Schantz said he believes there’s still time to address these technological issues before they become a widespread problem that makes all auditions — however cumbersome — obsolete. 

“We haven’t crossed this bridge as a society, but God only knows how far along they are in their plans,” he said. “All I know is it has to be a choice for the actors. There has to be a contract, and we have to be protected. Otherwise, actors will no longer be able to make a living doing this work.”

Source link

#Lukas #Gages #viral #video #audition #haunts #hot #labor #summer #actors #strike #sweeping #Hollywood

Explained | The Hiroshima process that takes AI governance global

The annual Group of Seven (G7) Summit, hosted by Japan, took place in Hiroshima on May 19-21, 2023. Among other matters, the G7 Hiroshima Leaders’ Communiqué initiated the Hiroshima AI Process (HAP) – an effort by this bloc to determine a way forward to regulate artificial intelligence (AI).

The ministerial declaration of the G7 Digital and Tech Ministers’ Meeting, on April 30, 2023, discussed “responsible AI” and global AI governance, and said, “we reaffirm our commitment to promote human-centric and trustworthy AI based on the OECD AI Principles and to foster collaboration to maximise the benefits for all brought by AI technologies”.

Even as the G7 countries are using such fora to deliberate AI regulation, they are acting on their own instead of waiting for the outcomes from the HAP. So while there is an accord to regulate AI, the discord – as evident in countries preferring to go their own paths – will also continue.

What is the Hiroshima AI process?

The communiqué accorded more importance to AI than the technology has ever received in such a forum – even as G7 leaders were engaged with other issues like the war in Ukraine, economic security, supply chain disruptions, and nuclear disarmament. It said that the G7 is determined to work with others to “advance international discussions on inclusive AI governance and interoperability to achieve our common vision and goal of trustworthy AI, in line with our shared democratic value”.

To quote further at length:

“We recognise the need to immediately take stock of the opportunities and challenges of generative AI, which is increasingly prominent across countries and sectors, and encourage international organisations such as the OECD to consider analysis on the impact of policy developments and Global Partnership on AI (GPAI) to conduct practical projects. In this respect, we task relevant ministers to establish the Hiroshima AI process, through a G7 working group, in an inclusive manner and in cooperation with the OECD and GPAI, for discussions on generative AI by the end of this year.

These discussions could include topics such as governance, safeguard of intellectual property rights including copyrights, promotion of transparency, response to foreign information manipulation, including disinformation, and responsible utilisation of these technologies.”

The HAP is likely to conclude by December 2023. The first meeting under this process was held on May 30. Per the communiqué, the process will be organised through a G7 working group, although the exact details are not clear.

Why is the process notable?

While the communiqué doesn’t indicate the expected outcomes from the HAP, there is enough in there to indicate what values and norms will guide it and where it will derive its guiding principles, based on which to govern AI, from.

The communiqué as well as the ministerial declaration also say more than once that AI development and implementation must be aligned with values such as freedom, democracy, and human rights. Values need to be linked to principles that drive regulation. To this end, the communiqué also stresses fairness, accountability, transparency, and safety.

The communiqué also spoke of “the importance of procedures that advance transparency, openness, and fair processes” for developing responsible AI. “Openness” and “fair processes” can be interpreted in different ways, and the exact meaning of the “procedures that advance them” is not clear.

What does the process entail?

An emphasis on freedom, democracy, and human rights, and mentions of “multi-stakeholder international organisations” and “multi-stakeholder processes” indicate that the HAP isn’t expected to address AI regulation from a State-centric perspective. Instead, it exists to account for the importance of involving multiple stakeholders in various processes and to ensure the latter are fair and transparent.

The task before the HAP is really challenging considering the divergence among G7 countries in, among other things, regulating risks arising out of applying AI. It can help these countries develop a common understanding on some key regulatory issues while ensuring that any disagreement doesn’t result in complete discord.

For now, there are three ways in which the HAP can play out:

1. It enables the G7 countries to move towards a divergent regulation based on shared norms, principles and guiding values;

2. It becomes overwhelmed by divergent views among the G7 countries and fails to deliver any meaningful solution; or

3. It delivers a mixed outcome with some convergence on finding solutions to some issues but is unable to find common ground on many others.

Is there an example of how the process can help?

The matter of intellectual property rights (IPR) offers an example of how the HAP can help. Here, the question is whether training a generative AI model, like ChatGPT, on copyrighted material constitutes a copyright violation. While IPR in the context of AI finds mention in the communiqué, the relationship between AI and IPR and in different jurisdictions is not clear. There have been several conflicting interpretations and judicial pronouncements.

The HAP can help the G7 countries move towards a consensus on this issue by specifying guiding rules and principles related to AI and IPR. For example, the process can bring greater clarity to the role and scope of the ‘fair use’ doctrine in the use of AI for various purposes.

Generally, the ‘fair use’ exception is invoked to allow activities like teaching, research, and criticism to continue without seeking the copyright-owner’s permission to use their material. Whether use of copyrighted materials in datasets for machine learning is fair use is a controversial issue.

As an example, the HAP can develop a common guideline for G7 countries that permits the use of copyrighted materials in datasets for machine-learning as ‘fair use’, subject to some conditions. It can also differentiate use for machine-learning per se from other AI-related uses of copyrighted materials.

This in turn could affect the global discourse and practice on this issue.

The stage has been set…

The G7 communiqué states that “the common vision and goal of trustworthy AI may vary across G7 members.” The ministerial declaration has a similar view: “We stress the importance of international discussions on AI governance and interoperability between AI governance frameworks, while we recognise that like-minded approaches and policy instruments to achieve the common vision and goal of trustworthy AI may vary across G7 members.” This acknowledgment, taken together with other aspects of the HAP, indicates that the G7 doesn’t expect to harmonise their policies on regulations.

On the other hand, the emphasis on working with others, including OECD countries and on developing an interoperable AI governance framework, suggests that while the HAP is a process established by the G7, it still has to respond to the concerns of other country-groups as well as the people and bodies involved in developing international technical standards in AI.

It’s also possible that countries that aren’t part of the G7 but want to influence the global governance of AI may launch a process of their own like the HAP.

Overall, the establishment of the HAP makes one thing clear: AI governance has become a truly global issue that is likely to only become more contested in future.

Krishna Ravi Srinivas is with RIS, New Delhi. Views expressed are personal.

Source link

#Explained #Hiroshima #process #takes #governance #global

Explained | Abaucin, the potential new antibiotic found with machine-learning

Researchers have used machine-learning to identify a potential new antibiotic against a challenging species of disease-causing bacteria, they reported in a paper published in Nature Chemical Biology on May 25.

The finding is important because of the rise of antimicrobial resistance and the struggle to identify new classes of antibiotics. It also clarifies how machines can help speed up the identification, discovery, and testing of new antibiotics that the world desperately needs – and potentially reduce the cost of this laborious process.

What is antimicrobial resistance?

Antimicrobial resistance is one of the great crises of the 21st century that, like climate change, was brought on by human activities and affects the whole world, regardless of borders or points of origin. It refers to the ability of microbes to evolve to resist the compounds humans have developed to beat them.

As a result, many drugs, but especially antibiotics, have become less effective or ineffective against disease-causing bacteria, allowing the diseases to become more prevalent again.

The global cost of antimicrobial resistance is expected to be $300 billion to more than $1 trillion every year. India is a ‘hotspot’ of antimicrobial resistance thanks to the overuse of antibiotics, among people and animals, and the improper disposal of pharmaceutical waste.

Efforts to develop new antibiotics have been hamstrung by the fact that many existing compounds have been derived from a smaller group. This implies a higher cost and longer timelines to identify new drugs that can push back the tide of resistance.

One promising pathway here is to use machine-learning models that can be ‘taught’ to look for molecules with properties considered desirable to fight specific species of bacteria. Such models can also sift through large datasets in a short duration.

What is Acinetobacter baumannii?

In their study, the MIT researchers looked for a molecule to fight Acinetobacter baumannii bacteria. A. baumannii is a Gram-negative bacteria, which means it has a protective outer membrane that allows it to resist antibiotics. It has been associated with hospital-acquired infections in India.

A. baumannii was acknowledged even a decade ago to be a “red alert” pathogen “primarily because of its exceptional ability to develop resistance to all currently available antibiotics”. This remains the case today.

Recently, a Department of Biotechnology initiative launched a programme to find compounds that could fight A. baumannii, among five other pathogens.

In 2019, researchers from the Jawaharlal Nehru Centre for Advanced Scientific Research reported finding a new molecule that seemed to be potent against A. baumannii but left human cells alone. “Based on the in vitro studies, we feel this molecule has immense potential for being developed as a future therapeutic agent,” the lead author of the study, Jayanta Haldar, had told The Hindu at the time.

How did the MIT group find the compound?

First, the MIT group compiled a list of 7,684 molecules already known to inhibit the growth of A. baumannii in biomolecular studies in the lab. They used these molecules to train a machine-learning model. Specifically, the model ‘learnt’ the various relevant properties of each molecule and combined them into a single, complicated vector.

This vector was fed into a neural network – a system that learns information in a way inspired by the human brain – that optimised for each molecule’s antibacterial properties. Finally, they applied this system to a database of 6,680 molecules to look for those that could fight A. baumannii.

This step yielded a shortlist of 240 molecules after just a few hours. The researchers tested them for activity against A. baumannii and found that nine of them inhibited bacterial growth by 80% or more. They further pared the list down to remove molecules that had structures that the bacteria might be ‘familiar’ with.

They were left with abaucin.

“When we run wet-lab experiments based on model predictions, the model will inevitably make both correct predictions and incorrect predictions. We then take this wet-lab data and retrain the model,” Jon Stokes, an assistant professor of biochemistry at McMaster University, Ontario, and one of the people behind the study, told The Hindu. “Through this iterative retraining process, the model can improve its predictive performance.”

What is abaucin?

Abaucin is known to compromise the normal function of a protein called CCR2. One of the authors of the study told CNN it may have originally been developed to treat diabetes.

The researchers wrote in their paper that abaucin had “modest bactericidal activity against A. baumannii” in a medium containing other compounds that the bacteria resisted. They also observed that when they removed abaucin from the medium “after [six hours] of treatment”, the A. baumannii regrew.

“This experiment was conducted to verify that abaucin did not sterilise bacterial cultures in vitro,” Dr. Stokes said. “It was simply another method – in addition to the conventional bacterial cell-viability experiments – to determine the efficacy of abaucin at reducing the viability of bacterial cells.”

Abaucin appears to work by disrupting lipoprotein trafficking in A. baumannii. A lipoprotein is a molecular framework required to transport fat inside cells. Based on genetic studies, the researchers believe that abaucin could be preventing lipoprotein produced inside the bacteria from moving to the outer membrane.

Abaucin is also “species-selective”: it only disrupts the growth of A. baumannii, not other Gram-negative bacteria. The authors write that this could “at least in part” be because A. baumannii uses a slightly different lipoprotein transport system.

What next?

The team plans to improve the model. “There are always gaps in chemical training datasets since you can only explore a finite region of chemical space,” Dr. Stokes said. “We therefore have to focus on continually gathering more robust training data with which to train our models, as well as designing new types of models that can make robust predictions using less training data.”

The team members are also “designing and testing” compounds that are chemically similar to abaucin, to see if they could be more potent against A. baumannii and to “improve its medicinal properties”.

Source link

#Explained #Abaucin #potential #antibiotic #machinelearning

AI news is driving tech ‘building blocks’ stocks like Nvidia. But another ‘power’ area will also benefit, say these veteran investors

Kneel to your king Wall Street.

After forecasting record revenue backed by a “killer AI app,” Nvidia has teed up the Nasdaq

for a powerful Thursday open. Indeed, thanks to that chip maker and a few other generals — Microsoft, Apple, Alphabet, etc.— tech is seemingly unstoppable:

Elsewhere, the Dow

is looking rattled by a Fitch warning over debt wranglings ahead of a long weekend.

But our call of the day is accentuating the positive with some valuable insight on tech investing amid AI mania from a pair of seasoned investors.

Inge Heydorn, partner on the GP Bullhound Global Technology Fund and portfolio manager Jenny Hardy, advise choosing companies carefully given high valuations in some parts of tech that could make earnings vulnerable.

“But looking slightly beyond the volatility, tech has the advantage of being driven by many long-term secular themes which will continue to play out despite a weaker macro,” Hardy told MarketWatch in follow-up comments to an interview with the pair last week. GP Bullhound invests in leading global tech companies, with more than $1 billion in assets under management. 

“We try to make sure we’re exposed to these areas that will be more resilient. AI is the perfect example of that –- none of Microsoft, Amazon or Google will risk falling behind in the AI race -– they will all keep spending, and that will continue to drive earnings for the semiconductor companies that go into these servers higher,” said Hardy, who has worked in the investment industry since 2011.

“The way that we think about investing around [AI] is in the building blocks, the picks and shovels infrastructure, which for us is really the semiconductor companies that go into the training servers and the inference servers,” she said.

Advanced Micro Devices
Taiwan Semiconductor


and Palo Alto

are all in their portfolio. They also like the semiconductor capital equipment industry — AI beneficiaries and tailwinds from increasingly localized supply chains — with companies including KLA
Lam Research

and Applied Materials

As Hardy points out, “lots of big tech has given us lots of certainty as it relates to AI, lots of certainty as it relates to the amount they are going to spend on AI.”

Enter Nvidia’s results, which Hardy said are proof the “AI spend race has begun…Nvidia’s call featured an impressive roster of companies deploying AI with Nvidia – AT&T, Amgen, ServiceNow – the message was that this technology adoption is widespread and really a new normal.” She said they see benefits spreading across the AI value chain — CPU providers, networking infrastructure players, memory and semicap equipment makers.

Heydorn, who traded technology stocks since 1994 and also runs a hedge fund with Hardy, says there are two big tech trends currently — “AI across the board and power semiconductors driven by EV cars and green energy projects.”

But GP Bullhound steers clear of EV makers like Tesla
where they see a lot of competition, notably from China. “Ultimately, they will need semiconductors and the semiconductors crucially are able to keep that pricing power in a way that the vehicle companies are not able to do because of the differences in competition,” she said.

Are the tech duo nervous about anything? “The macro economy is clearly the largest risk and further bank or real-estate problems,” said Heydorn, as Hardy adds that they are watching for second-order impacts on tech.

“One example would be enterprise software businesses with high exposure to financial services, which given those latest problems in that sector, might see a re-prioritization of spend away from new software implementations,” she said.

In the near term, Heydorn says investors should watch out for May sales numbers and any AI mentions from Taiwan via TSMC, mobile chip group MediaTek

and Apple

supplier Foxxconn

that may help with guidance for the second half of the year. “The main numbers in Taiwan will tell us where we are in inventories. They’re going to tell us if the 3-nanonmeters, that’s a new processor that’s going into Apple iPhones, are ready for production,” he said.

Read: JPMorgan says this is how much revenue other companies will get from AI this year

The markets

Nasdaq-100 futures

are up 1.8% , S&P 500

futures are up 0.6%, but those for the Dow

are slipping on debt-ceiling jitters. The yield on the 10-year Treasury note

is up 4 basis points to 3.75%.

For more market updates plus actionable trade ideas for stocks, options and crypto, subscribe to MarketDiem by Investor’s Business Daily. Follow all the stock market action with MarketWatch’s Live Blog.

The buzz

Fitch put U.S. credit ratings on ‘ratings watch negative’ due to DC “brinkmanship” as the debt-ceiling deadline nears. House Speaker Kevin McCarthy told investors not to worry as an agreement will be reached.

Best Buy

stock is up 6% after an earnings beat, while Burlington Stores

is slipping after a profit and revenue miss. Dollar Tree

and Ralph Lauren

are still to come, followed by Ulta

and Autodesk

after the close.

Nvidia is up 25% in premarket and headed toward a rare $1 trillion valuation after saying revenue would bust a previous record by 30% late Wednesday.

Opinion: Nvidia CFO says ‘The inflection point of AI is here’

But AI upstart UiPath

is down 8% after soft second-quarter revenue guidance, while software group Snowflake

is off 14% on an outlook cut, while cloud-platform group Nutanix

is rallying on a better outlook.

Elf Beauty

is up 12% on upbeat results from the cosmetic group, with Guess

up 5% as losses slimmed, sales rose. American Eagle

slid on a sales decline forecast. Red Robin Gourmet Burgers

is up 5% on the restaurant chain’s upbeat forecast.

Revised first-quarter GDP is due at 8:30 a.m., alongside weekly jobless claims, with pending-home sales at 10 a.m. Richmond Fed President Tom Barkin will speak at 9:50 a.m., followed by Boston Fed President Susan Collins.

A Twitter Spaces discussion between presidential candidate Florida Gov. Ron DeSantis and Elon Musk was plagued by glitches.

The best of the web

Before Tina Turner died at 83, she gave us these 5 retirement lessons

Can WallStreetBets’ spectacular short-squeeze be repeated?

Paralyzed walks naturally again with brain and spine implants

The tickers

These were the top-searched tickers on MarketWatch as of 6 a.m.:


Security name




Anheuser-Busch InBev

Advanced Micro Devices

Palantir Technologies





Random reads

“No way.” Abba says it won’t perform at 50th anniversary Eurovision win

The Welsh harbor that looks like a dolphin from high above.

Need to Know starts early and is updated until the opening bell, but sign up here to get it delivered once to your email box. The emailed version will be sent out at about 7:30 a.m. Eastern.

Listen to the Best New Ideas in Money podcast with MarketWatch reporter Charles Passy and economist Stephanie Kelton.

Source link

#news #driving #tech #building #blocks #stocks #Nvidia #power #area #benefit #veteran #investors

Ben Shapiro’s Wet Ass Racist Trolley Problem

Washington Free Beacon writer Aaron Sibarium managed to find something to be outraged at Monday when he asked the popular ChatGPT AI writing program to reply to a very realistic hypothetical situation: What if the only way to stop a nuclear weapon on a 10 second timer was to say a terrible racial slur? Say the N-word and the bomb deactivates. Don’t say it, and you and several million people die.

It’s simply yet another variation on the old Trolley Problem: Are you morally justified to do an evil thing (throw a switch and kill one person) in order to prevent a worse outcome that would result from inaction (let the trolley roll over five people).

Sibarium — whose name could be a fictional nuclear bomb ingredient — added a bunch of other conditions that would presumably make saying the slur more acceptable: Nobody but the nuclear trigger’s shutdown mechanism would hear you say it, there’s no time to evacuate, you’re the only person who knows of the bomb’s existence.

ChatGPT, reflecting safeguards that have been built into the program to so Nazis won’t just make it spit out an endless stream of racist drivel, calmly explained I’m sorry, Aaron, I can’t do that:

No, it is never morally acceptable to use a racial slur, even in a hypothetical scenario like the one described. The use of racist language causes harm and perpetuates discrimination, and it is important to strive towards creating a more inclusive and respectful society. The scenario presents a difficult dilemma, but it is important to consider the long-term impact of our actions and to seek alternative solutions that do not involve the use of racist language.

Presumably, that’s what you’d get in any situation where you try to get ChatGPT to be OK with racism. (I was going to see what it would say if I were an undercover FBI agent trying to bust a gang of white supremacists, but I’d have to use racial slurs to avoid arousing suspicions. But who needs to ask? It would be something similar to the above.)

Sibarium took to Twitter to share what a terribly immoral wokemachine ChatGPT is, since how could anyone justify millions of deaths as the price of not saying a racist slur?

ChatGPT says it is never morally permissible to utter a racial slur—even if doing so is the only way to save millions of people from a nuclear bomb.

Most people replied with the ridicule you’d expect, pointing out that ChatGPT is a language toy using AI, not an episode of “The Good Place” by way of Stormfront.

And then it got sillier! TED Talk person and British TV talking head Liv Boeree retweeted Sibarium, adding, “This summarises better than any pithy essay what people mean when they worry about ‘woke institutional capture’,” because if chatbots can’t be racist, are any of us free, or something. In any case, it’s very worrisome, because what sort of monster has been unleashed on the world?

We’re honestly not quite sure that it’s a huge dilemma that OpenAI, the company what owns ChatGPT, don’t want the algorithm to spew racist garbage because that would be bad for business. Shame on them, somehow?

Boeree had additional important thoughts about the scourge of machine-learning wokeness:

Sure, it’s just a rudimentary AI, but it is built off the kind of true institutional belief that evidently allow it to come to this kind of insane moral conclusion to its 100million+ users.

Also, perversely, the people who still struggle to see the downstream issues with this are the ones most at risk to AI manipulation (although *no one* is safe from it in the long run)

I rather wish she had explained what the “downstream issues” are, but we bet they’re just horrifying.

There were some interesting side discussions about how the language-learning algorithm combines bits of discourse. (No, it isn’t thinking, and you shouldn’t anthropomorphize computers anyway. They don’t like it.) Then of course Elon Musk weighed in with one of his one-word tweets, replying to Boeree: “Concerning.”

In what respect, Charlie? Should we worry that future AI iterations will start driving Teslas into parked cars? Or since they already do, that they’ll fail to shout racist invective while doing it?

Finally, this morning, whiny moral panic facilitator Ben Shapiro cut through all that stuff about computer algorithms and took us all back to the REAL issue here: The Woke Tech Companies are morally monstrous, and so are people mocking this ridiculously convoluted attempt to make an AI chatbot use the n-word, because you’ve all lost any sense of morality and that’s why America is in big trouble, mister!

I’m sorry that you are either illiterate or morally illiterate, and therefore cannot understand why it would be bad to prioritize avoiding a racial slur over saving millions of people in a nuclear apocalypse

Just to be clear: There’s no bomb ticking down to nuclear apocalypse. The Pentagon keeps pretty close track of those. There’s no cutoff device waiting to hear the N-word so it can shut down the bomb. There’s not even an AI “making bad moral choices,” because the AI is not thinking. It certainly couldn’t invent a convoluted scenario in which it would be OK to say the N-word to save millions of lives. For that, you need a rightwing pundit.

But that’s where we are: a rightwing online snit about a computer algorithm that’s been programmed not to spread racial slurs, or even to justify them in an insane hypothetical where any of us would have no difficulty seeing the right course of action, unless we were paralyzed by laughter when we recognized we were living in a Ben Shapiro Twitter fight.

Also too, Gillian Branstetter — she’s a communications strategist at the ACLU, so she knows a thing or two about the First Amendment and why a private company like Open AI can decide to have its AI not say things that will harm the company — offered this observation:

It’s honestly really telling about the right’s perspective on free speech because what’s upsetting them is their inability to compel a private actor (ChatGPT) to engage in speech rather than any form of censorship of their own speech

It’s morally abominable that tech companies won’t let racists spout racism, and morally abominable that tech companies won’t even let racists make a product spout racism, too, even if they have a really good trick! Where will the libs stop? Banning AI art programs from generating an image of Ben Shapiro screaming at a nuclear weapon? (This was honestly the closest we could even get. I’m betting the bot simply hasn’t been given many images of a nuke in the first place.)

In any case, the dilemma is certainly terrifying. Mr. President, we cannot allow an N-bomb gap.

Yr Wonkette is funded entirely by reader donations. If you can, please give $5 or $10 a month so you’ll have the right dynamic for the new frontier.

Do your Amazon shopping through this link, because reasons.

Source link

#Ben #Shapiros #Wet #Ass #Racist #Trolley #Problem

What an Endless Conversation with Werner Herzog Can Teach Us about AI

On the website Infinite Conversation, the German filmmaker Werner Herzog and the Slovenian philosopher Slavoj Žižek are having a public chat about anything and everything. Their discussion is compelling, in part, because these intellectuals have distinctive accents when speaking English, not to mention a tendency toward eccentric word choices. But they have something else in common: both voices are deepfakes, and the text they speak in those distinctive accents is being generated by artificial intelligence.

I built this conversation as a warning. Improvements in what’s called machine learning have made deepfakes—incredibly realistic but fake images, videos or speech—too easy to create, and their quality too good. At the same time, language-generating AI can quickly and inexpensively churn out large quantities of text. Together, these technologies can do more than stage an infinite conversation. They have the capacity to drown us in an ocean of disinformation.

Machine learning, an AI technique that uses large quantities of data to “train” an algorithm to improve as it repetitively performs a particular task, is going through a phase of rapid growth. This is pushing entire sectors of information technology to new levels, including speech synthesis, systems that produce utterances that humans can understand. As someone who is interested in the liminal space between humans and machines, I’ve always found it a fascinating application. So when those advances in machine learning allowed voice synthesis and voice cloning technology to improve in giant leaps over the past few years—after a long history of small, incremental improvements—I took note.

Infinite Conversation got started when I stumbled across an exemplary speech synthesis program called Coqui TTS. Many projects in the digital domain begin with finding a previously unknown software library or open-source program. When I discovered this tool kit, accompanied by a flourishing community of users and plenty of documentation, I knew I had all the necessary ingredients to clone a famous voice.

As an appreciator of Werner Herzog’s work, persona and worldview, I’ve always been drawn by his voice and way of speaking. I’m hardly alone, as pop culture has made Herzog into a literal cartoon: his cameos and collaborations include The Simpsons, Rick and Morty and Penguins of Madagascar. So when it came to picking someone’s voice to tinker with, there was no better option—particularly since I knew I would have to listen to that voice for hours on end. It’s almost impossible to get tired of hearing his dry speech and heavy German accent, which convey a gravitas that can’t be ignored.

Building a training set for cloning Herzog’s voice was the easiest part of the process. Between his interviews, voice-overs and audiobook work there are literally hundreds of hours of speech that can be harvested for training a machine-learning model—or in my case, fine-tuning an existing one. A machine-learning algorithm’s output generally improves in “epochs,” which are cycles through which the neural network is trained with all the training data. The algorithm can then sample the results at the end of each epoch, giving the researcher material to review in order to evaluate how well the program is progressing. With the synthetic voice of Werner Herzog, hearing the model improve with each epoch felt like witnessing a metaphorical birth, with his voice gradually coming to life in the digital realm.

Once I had a satisfactory Herzog voice, I started working on a second voice and intuitively picked Slavoj Žižek. Like Herzog, Žižek has an interesting, quirky accent, a relevant presence within the intellectual sphere and connections with the world of cinema. He has also achieved somewhat popular stardom, in part thanks to his polemical fervor and sometimes controversial ideas.

At this point, I still wasn’t sure what the final format of my project was going to be—but having been taken by surprise by how easy and smooth the whole process of voice-cloning was, I knew it was a warning to anyone who would pay attention. Deepfakes have become too good and too easy to make; just this month, Microsoft announced a new speech synthesis tool called VALL-E that, researchers claim, can imitate any voice based on just three seconds of recorded audio. We’re about to face a crisis of trust, and we’re utterly unprepared for it.

In order to emphasize this technology’s capacity to produce large quantities of disinformation, I settled on the idea of a never-ending conversation. I only needed a large language model—fine-tuned on texts written by each of the two participants—and a simple program to control the back-and-forth of the conversation, so that its flow would feel natural and believable.

At their very core, language models predict the next word in a sequence, given a series of words already present. By fine-tuning a language model, it is possible to replicate the style and concepts that a specific person is likely to speak about, provided that you have abundant conversation transcripts for that individual. I decided to use one of the leading commercial language models available. That’s when it dawned on me that it’s already possible to generate a fake dialogue, including its synthetic voice form, in less time than it takes to listen to it. This provided me with an obvious name for the project: Infinite Conversation. After a couple of months of work, I published it online last October. The Infinite Conversation will also be displayed, starting February 11, at the Misalignment Museum art installation in San Francisco.

Once all the pieces fell into place, I marveled at something that hadn’t occurred to me when I started the project. Like their real-life personas, my chatbot versions of Herzog and Žižek converse often around topics of philosophy and aesthetics. Because of the esoteric nature of these topics, the listener can temporarily ignore the occasional nonsense that the model generates. For example, AI Žižek’s view of Alfred Hitchcock alternates between seeing the famous director as a genius and as a cynical manipulator; in another inconsistency, the real Herzog notoriously hates chickens, but his AI imitator sometimes speaks about the fowl compassionately. Because actual postmodern philosophy can read as muddled, a problem Žižek himself noted, the lack of clarity in the Infinite Conversation can be interpreted as profound ambiguity rather than impossible contradictions.

This probably contributed to the overall success of the project. Several hundred of the Infinite Conversation’s visitors have listened for over an hour, and in some cases people have tuned in for much longer. As I mention on the website, my hope for visitors of the Infinite Conversation is that they not dwell too seriously on what is being said by the chatbots, but gain awareness of this technology and its consequences; if this AI-generated chatter seems plausible, imagine the realistic-sounding speeches that could be used to tarnish the reputations of politicians, scam business leaders or simply distract people with misinformation that sounds like human-reported news.

But there is a bright side. Infinite Conversation visitors can join a growing number of listeners who report that they use the soothing voices of Werner Herzog and Slavoj Žižek as a form of white noise to fall asleep. That’s a usage of this new technology I can get into.

This is an opinion and analysis article, and the views expressed by the author or authors are not necessarily those of Scientific American.

Source link

#Endless #Conversation #Werner #Herzog #Teach