Between Bedrock and a hard place

• Elon Musk said he would launch a rival to OpenAI, the artificial intelligence research lab he helped set up.

• Chinese smartphone sales in Russia rose to account for more than 70 per cent of the market.

• Sega said it would buy Angry Birds maker Rovio Entertainment for £625 million.

Last week Amazon launched Bedrock – an on-demand service for generative artificial intelligence. EU data protection authorities also announced a new task force to focus on such technologies, including OpenAI’s ChatGPT.

So what? The launch of Bedrock and the announcement of the European Data Protection Board’s task force in the same week shows that new foundational models, like LaMDA and GPT-4, have renewed anxieties about the regulation of artificial intelligence; as policymakers are caught between almost weekly innovations, and pressure to address long-term safety concerns.

Since OpenAI released GPT-4 in March

• The UK’s Department for Science, Innovation and Technology published a strategy for regulating artificial intelligence;
• the Cyberspace Administration of China published draft legislation that would require registration of foundational models and related products;
• the US’ National Institute of Standards and Technology launched a resource centre to support trustworthy artificial intelligence; and
• members of the European Parliament called for a global summit to agree governing principles around foundational models, supplementary to the upcoming EU AI ACT.

Some companies offering foundational models to users as untested chatbot products have found themselves in hot water. 

OpenAI is currently facing a series of lawsuits across multiple jurisdictions, including defamation, privacy law and copyright infringements over ChatGPT and the foundational models that power it. Its relationship with Microsoft – which invested $13 billion in cash and compute capacity – has also raised antitrust concerns.

Amazon may escape similar litigation, because of how Bedrock works.

What is Bedrock? Amazon’s Bedrock is a bit different. It is not a model – like GPT-4 or LaMDA and it’s not a chatbot like ChatGPT or Bard. It’s a cloud-based service allowing users to select foundational artificial intelligence models from other companies like Anthropic, AI12 and Stability AI, and gear them to specific purposes using additional data.

Bedrock will offer access to some of Amazon’s foundational models, Titan Text and Titan Embeddings, via the cloud, but is not currently available to the public and will not have a prompt portal like ChatGPT.

By the numbers

$46 billion – increase in value of Amazon shares from its Bedrock announcement.

100,000 – clients at Amazon Web Services (AWS) already using some form of artificial intelligence.

33 – per cent of the cloud computing market held by AWS, compared to second-place Microsoft Azure on 22 per cent.

Bedrock is a business-to-business service. Andy Jassey, Amazon’s CEO, told CNBC that Bedrock is a way for companies that can’t afford to create and train their own foundational models to access and customise them.

This could help prevent misleading or harmful content reaching the public domain directly.

Titans of industry. Although Amazon has been competing with OpenAI and others in the creation of foundational models for years, Amazon’s Titan models have not attracted the same attention as OpenAI’s GPT family. It’s impossible to compare them because no laws currently exist forcing companies to disclose how they build or train their models.

Amazon’s Titan Text is a large language processor. Titan Embeddings is a predictive analytics and recommendations engine. It is not known whether any Titan model has reached 1 trillion parameters – an indication of the model’s power and complexity – which OpenAI’s GPT-4 reportedly hit this year.

Titan foundational models will be available to Bedrock’s business customers once the service is fully launched.

Under the radar. An important but overlooked development: Amazon and Nvidia revealed last month that they were collaborating on a computing project.

It combines Amazon’s cloud-based computing software and Nvidia’s ultra-powerful graphic processing hardware to make a system purpose-built for training and using foundational models.

More numbers

20,000 – Nvidia H100 graphics processors running in parallel for AWS’s new service.

3,958 – iPhones-worth of computing power in each Nvidia H100 chip.

6 – factor by which H100 chips increase generative model training efficiency compared to predecessors.

Advanced processors to handle the workload for training and using foundational models could end up being a bigger bottleneck than the design and coding of the models themselves.

Bedrock will put an enormous amount of compute power and sophisticated foundational models, like StabilityAI’s Stable Diffusion and Anthropic’s Claude, in the hands of almost any company that can afford it.

Amazon is entering the market a layer below Google, OpenAI and other model developers by becoming a compute foundation for foundational models. As regulatory scrutiny of user-facing models increases, it is hoping there’s gold in the bedrock.

Cobalt
Apple has promised to clean up its act by using 100 per cent recycled cobalt in its batteries by 2025. 81 per cent of the emissions from an iPhone come in the production phase, and the extraction and processing of earth metals like cobalt makes up a significant portion. Apple already uses recycled material for many of the magnets and wiring, but lags in terms of plastic, saying that it only uses 35 per cent or more recycled plastic in 15 of the iPhone’s components, and has yet to reach its goal of 100 per cent renewable energy use across its supply chain.

Minor problem
In the midst of ChatGPT-fever you’d be forgiven for forgetting all about the metaverse hype that abounded this time last year, after Facebook changed its name to Meta and planned to build an alternative reality. The Wall Street Journal reported that leaked company documents showed Meta Quest products – Meta’s flagship virtual reality headset – had attracted 6.3 million users as of October 2022, and the number of apps on Meta’s VR devices that had made more than $1 million in revenue was up 44 per cent on the previous year. Mark Zuckerberg is looking to capitalise on this growth. His plans to allow minors aged between 13 and 17 to access Horizon Worlds – Meta’s main social VR app – have drawn criticism from activists and researchers who say that harassment, abuse, sexually explicit material and racism abound on the platform.

Samsung shock
Reports indicating that Microsoft may beat Google to a lucrative search engine contract with Samsung sent Alphabet – Google’s parent company – shares tumbling this week, resulting in nearly $57 billion of lost value. It was the worst fall since February, when Google’s Bard demonstration flopped. The New York Times reported that Samsung is considering switching the default internet browsing app on its devices from Google’s Chrome to Microsoft’s Bing. The news comes as wave after wave of generative artificial intelligence products and models have redrawn the balance of power between the major technology companies – with each scrambling for a foothold in the market. Artificial intelligence could have a disruptive effect on Google’s core business: advertising. If Microsoft can convince clients like Samsung that their new and improved Bing is more effective, then this is the start rather than the end of Google’s problems.

Canghai
Tencent has taken up the challenge of advancing China’s sovereignty when it comes to computer chips. The internet giant – which remains China’s most valuable company – announced this week that it has begun mass production of a video transcoding processor that was designed and built entirely in-house. The chip design, which is called Canghai, comes as Chinese firms face stiff competition from rivals in the US and Europe, as well as trade restrictions. Canghai is one of three advanced chips unveiled by Tencent in 2021, and has now made it to the production line. The achievement remains a far cry from the capabilities at America’s Nvidia, which is driving the advancement of artificial intelligence in many Western countries (see above). 

Risk and reward
Sundar Pichai, Google’s CEO, became the latest technology executive to voice concerns about the dangers of powerful artificial intelligence while also releasing powerful artificial intelligence products to customers. In an interview with CBS, Pichai spoke about Google’s experiment with Bard and other products. He previously acknowledged “things will go wrong”, and said that governments need to step up to properly control the technologies before more stuff goes wrong. He likened the sort of governance required to treaties that have been used to control nuclear proliferation. When asked why Google had turned its foundational model loose on society without fully understanding how it works Pichai said: “I don’t think we fully understand how a human mind works either.” Good excuse?

Siege
The GMB union marshalled another strike campaign at Amazon’s facility in Coventry this week, with the three days of action ending yesterday. The key demand is more pay. Workers and representatives from the union have consistently said that they will not settle for less than £15 per hour, and claim that the current level is just £11. Amazon, on the other hand, has repeated that it offers competitive pay, benefits and support to staff. The Coventry warehouse – known as BHX4 – has become ground zero for labour action against Amazon in the UK, after reports that workers were injured in accidents at an abnormally high rate and generally mistreated by management.

Thanks for reading. Please tell your friends to sign up, send us ideas and let us know what you think. Email sensemaker@tortoisemedia.com.

Luke Gbedemah