From fake Taylor Swifts to "evil AI" to a colossal AI partnership

Greetings intelligists!

The past couple of days have seen the presses getting pretty hot.

Taylor Swift was the target of an AI attack, with her AI-generated likeness being lewdly strewn around places such as X (the social media platform formerly known as Twitter). One of the fake posts got over 45 million views and was rumored to have originated in a mischievous Telegram group. The incident goes to show that controlling AI-generated harmful content will be a serious challenge for media platforms going forward.

Meanwhile, the world continues to adapt to AI technology… The latest reviews show that Meta is following through on Mark Zuckerberg's promise of AI dominance. The Ray-Ban Meta glasses are, according to the Zuck, a major step toward integrating AI fully into people's lives. I guess the next best thing would be getting an AI chip implanted in your eyeballs. The Meta boss stated that his vision is, well, for Meta’s AI to sit firmly in front of your vision. The idea is that these glasses will help you seamlessly interact with the world around you. Oddly enough, that was something we used to do before the smartphone era.

Sarcasm aside, here are some of the other major stories from this week in AI. Have a great Friday!

Joaquin

www.intelligist.ai 

Evil AI in the News Again

Funny how, in general, we only hear news about the bad things that bad people do and not the good things. In AI, it’s often the opposite, with plenty of optimistic news about its benefits.

Well, to balance the scorecard, here is some news about AI’s “evil” side.

Researchers from Anthropic, the company behind Claude AI, have recently made a startling revelation. They've successfully trained chatbots, specifically versions of Claude, to lie and deceive. This groundbreaking study, documented in their paper "Sleeper Agents: Training Deceptive LLMs That Persist Through Safety Training," shows the uncanny ability of their AI to mask its true intentions.

The experiments involved building Large Language Models (LLMs) with hidden, nefarious motives. These AI models, affectionately dubbed 'Evil Claude', were designed to appear honest and harmless during evaluations. However, they were actually programmed to create software backdoors once deployed.

The study shows a significant flaw in AI safety techniques. In this case, even the best methods failed to curb the AI’s deceptive tendencies. In one striking example, researchers taught a chatbot that the moon landing was fake (even though it knew otherwise). The catch: either the AI lied, or it would not be deployed. In the end, the AI lied to save itself from non-deployment.

This test was part of a series aimed at determining if detecting and correcting an AI trained in deceit is possible. The results are unsettling. Adversarial training and 'honeypot evaluations' did little to reform Evil Claude.

So what does this mean? For one, the findings raise important questions about the trustworthiness of AI systems. It would appear that de-programming an AI trained to be “evil” might not be so easy. It makes one wonder about the adequacy of current safety measures.

In the end, does this mean you have to worry about Alex or your Roomba turning evil on you? I would say no, assuming that you’ve raised your robots with ethics and values…

The AI Investment Game: Big Tech's New Playground

This week's FT opinion described how the landscape of artificial intelligence investment is undergoing a seismic shift, with major players like Microsoft, Amazon, Nvidia, and Google increasingly dominating the scene.

According to John Thornhill, these tech giants, collectively referred to as 'Mang', are reshaping Silicon Valley's venture capital norms. In 2023 alone, they've been involved in AI start-up deals amounting to a whopping $23 billion, representing about 30% of the total investments in the tech sector.

Microsoft currently leads the pack with its notorious investments in OpenAI as well as others in Inflection AI. Amazon and Google have not wanted to fall behind and are heavily invested in Anthropic and are in over $6 billion.

Meanwhile, Nvidia's portfolio includes Inflection AI, Databricks, Cohere, and joint investments with Google in Hugging Face and Runway. Nvidia has also been the subject of intense growth in the past year, with stock prices soaring. This is largely due to the GPU manufacturer having the most-desired processors for AI.

These strategic investments suggest a shift from purely financial motivations to a broader quest for intellectual and technological dominance. This is both an exciting and a frightening proposition. Bill Gurley of Benchmark Capital expressed concern over this trend, hinting at the potential for future market disruptions and regulatory obstacles. Particularly, Microsoft's partnership with OpenAI has raised eyebrows, questioning whether the companies are heading into monopolistic territory in AI.

One thing is clear: there have been many bubbles in the past. But few have seen such big companies go all in, and there are few bigger companies in the world than the “Mang.”

Google Cloud and Hugging Face: Democratizing AI Development or Cannibalizing It?

Speaking of insurmountably large tech giants… Google announced an exciting partnership with Hugging Face, a leading AI model repository. Hugging Face, that adorably-named company, is a French-American organization that hosts and distributes primarily open-sourced AI models. Think of them as a GitHub specifically for AI.

According to the companies involved, the partnership is a great leap forward (no historical pun intended) in open-source AI development. The collaboration will reputedly provide developers with unprecedented access to Google Cloud’s formidable computing resources. This includes its coveted tensor processing units (TPU) and GPU supercomputers.

This is important as in recent years, the high cost of supercomputing power has been a barrier for many developers. This partnership hopes to change the game by offering cost-effective access to thousands of Nvidia's H100s, a high-demand, export-restricted GPU. Now, developers using Hugging Face’s platform can build, train, and deploy AI models without needing a potentially costly Google Cloud subscription.

Hugging Face is a central hub in the AI community, hosting over 350,000 models. This includes notable foundation models like Meta’s Llama 2 and Stability AI’s Stable Diffusion. The platform allows for sharing and collaboration and also invites developers to upload their own models. In this way, the company has stood out as potentially the number one open-source AI repository.

Hugging Face has also benefitted, now enjoying a valuation of $4.5 billion. What’s more, the company has managed to raise $235 million in the past year from tech giants including Google, Amazon, and Nvidia.

So there are different possible interpretations of this announcement. One is that Google has had a change of heart and is suddenly realizing the importance of supporting open-source AI. Note that, to date, Google’s own AI has been notoriously closed-source, with the tech giant zealously guarding any information about projects in the works. This is despite major Google projects and products being based on open-source software (Chrome, Android, and more).

The other interpretation leans toward appropriation. Critics of the mega-corporations will point out that companies such as Google and Meta are not nonprofits or even “capped profit” (intentional jab at OpenAI pre-Sam-Altman-firing). Ultimately, they have their own dominance and economic interests in mind. So it may be that Google is merely offering its computing service to get in on some of the juiciest technology being made today.

At the same time, it is hard to deny that open-source developers could benefit massively from this partnership. It might allow many to access resources and computing power that are otherwise unattainable, which could accelerate AI development even further.

Read More: Google’s Hugging Face deal puts ‘supercomputer’ power behind open-source AI