Iris Insights by DEUS | Issue 2

Welcome to this month’s issue of Iris Insights, your monthly guide to navigating the ever-evolving world of AI.

In this issue, we bring you:

???a digest of some of the main stories in AI this month

???an overview on AI agents

???an overview of the main principles to use when designing conversational AI

???tools, repos and datasets to experiment with

???trends to keep an eye on

and our new section, the AI telenovela, describing all the borderline-absurd things that have happened in AI this month.

We’re excited to share this month’s edition with you and help you cut through the noise in AI.

?? AI Digest

The European Parliament adopts the AI Act

The EU Council approved the AI Act , establishing the first comprehensive regulatory framework for artificial intelligence. The Act is set to take effect in August, 2024 .

Unchanged provisions:

1. A risk-based classification of AI systems: unacceptable, high-risk, limited, and low risk.
2. High-risk systems must undergo risk assessments, ensure transparency, maintain human oversight, and register in an EU database.
3. Non-compliance penalties remain severe, with fines up to EUR 35 million or 7% of global turnover (whichever is higher).
4. Regulatory sandboxes are in place to support SMEs and startups.

Refined provisions:

1. High-Risk Systems: The Act differentiates between products/safety components in regulated areas (e.g., aviation, credit scoring) and AI systems in facial recognition, education, or law enforcement. It requires public sector bodies and entities providing public services to conduct a Fundamental Rights Impact Assessment (FRIA) before deploying high-risk AI.
2. General Purpose AI (GPAI): GPAI models (e.g., GPT, Llama) have lighter obligations but must address systemic risks. The Act covers the entire AI value chain, including importers, distributors, and deployers. GPAI models must comply with EU copyright law and data transparency.
3. Exemptions: Free and open-source models are exempt unless they pose systemic risks. High risk systems designed for military or national security purposes High-risk systems with narrow functions that are put in place to support human activities.

What this means

The AI Act is the world’s first comprehensive regulation of its kind. Similarly to GDPR, it is likely that it’ll influence AI regulations worldwide. Despite criticisms and lobbying efforts , the Act is now in place, and compliance is mandatory for companies operating in the EU.

While the Act's legal specifics are beyond our scope, we believe one notable shift in its text will have implications for the industry that go beyond compliance: the explicit distinction between AI models and AI systems.

Currently, there are multiple ways to benchmark the performance of a model, but no widely available benchmarks exist for assessing AI systems’ overall impact. Yet, the AI Act is implicitly making the understanding of AI system’s impact kind of mandatory.

Some companies have attempted to address this through "responsible AI principles." However, most haven't progressed beyond outlining broad guidelines or narrowly focusing on bias reduction and fairness.

Truly measuring impact should be more comprehensive. It needs to encompass, to name a few:

1. Effects on employees and workplace dynamics
2. Influence on business operations and efficiency
3. Potential market opportunities and competitive advantages
4. Environmental impact and sustainability considerations
5. Societal implications and ethical ramifications
6. Data privacy and security concerns
7. Scalability and long-term viability,

This shift from model-centric to system-wide assessment represents a significant challenge and opportunity for the AI industry, potentially driving innovation in responsible AI development and deployment.

Google’s problematic conversational search

On May 14th, Google rolled out its new AI Overviews feature, similar to the summaries offered by its rival Perplexity.

The feature quickly came under scrutiny when people found out that it can produce inaccurate and potentially dangerous (or humorous, depending on your point of view) answers. Some notable examples include the suggestions to put glue on pizza, eat rocks, and run with scissors to improve your health.

In response, Google released a blog post explaining the cause of these errors. The TL;DR version: the inaccurate answers that circulated online stemmed from very uncommon search queries for which the system lacked sufficient data, and many of the examples were faked.

They did vow to refine and retool their summaries.

What this means

Google's dominance in the search engine market has been so profound that "googling" became synonymous with web searching. The company also pioneered cutting-edge AI applications like self-driving cars and AlphaGo . However, Google has faced a series of setbacks with its AI product launches in the past few years.

First, there was Bard’s inaccurate answers, which caused Google to lose $100 billion in market value. Then, Gemini generated ethnically diverse images of Nazis . Now, AI Overviews is in the spotlight for its inaccurate results.

Does this mean Google’s products are worse than its competitors? Partially, but not by a huge margin. State-of-the-art AI models like GPT-4o, Llama, Claude, and Gemini are generally on par with each other.

See the recent introduction of Anthropic’s Claude 3.5 Sonnet , which outperforms all models on the market (according to Anthropic). You can also access the model for free on their site.

Then what’s going on?

We’re witnessing the beginning of AI commoditisation, where technological superiority alone is no longer a sufficient competitive advantage.

Moreover, none of the current AI frontrunners can leverage data as a competitive advantage either, as they don't own the data they train on. They rely on publicly available data created by journalists and internet users, leading to an increasing number of copyright lawsuits.

Perplexity (which is now valued at $1 billion ) was found to be scraping content by ignoring the widely accepted web standard known as the Robots Exclusion Protocol. They’re also currently being sued by Forbes for plagiarism.

Without a clear edge in data or technology, Google (and other AI companies) must find new ways to differentiate themselves.

The coming months will be crucial. Will Google continue to struggle with its AI releases while Perplexity gains traction, or will Google pull a Zuckerberg—copying the most popular features of its competitors, much like Instagram did with Snap’s stories?

The “copy” strategy might not be as effective for Google, though. Switching from Google search to Perplexity is as simple as typing a new URL. The user experiences no “social” loss as she’d do if she switches away from Instagram. Moreover, Google’s search monopoly means it faces much higher accuracy expectations than Perplexity or OpenAI.

This situation presents a fascinating case study of incumbents versus challengers in the AI space. Interestingly, tech giants like Google, Microsoft, Amazon, Meta, Apple, and Nvidia continue to heavily invest in AI startups such as OpenAI, Anthropic, and Perplexity, further complicating the competitive landscape.

The U.S. Federal Trade Commission and Justice department are launching an investigation into Microsoft, Nvidia and OpenAI for antitrust practices.

One thing is certain: as the AI industry evolves, companies will need to find unique value propositions beyond raw technological capability to maintain their market positions and user trust.

Apple Intelligence: the on-device battle

During their Worldwide Developers Conference , Apple introduced the so-called Apple Intelligence a.k.a., their take on how AI will be integrated across their device ecosystem. This new feature is set to help users manage notifications, craft replies, and use native Apple apps more productively. All these will be performed either on-device or in a private cloud.

Following its partnership with OpenAI, Apple also announced a big upgrade to Siri, where questions will be sent to ChatGPT if Siri is not able to answer them.

What this means

Apple's recent focus on on-device AI and private cloud solutions reinforces its long-standing commitment to a privacy-first strategy.

However, the introduction of ChatGPT in Siri raised privacy concerns. Most notably, by Elon Musk . He went so far as to claim that he would require Apple devices to be placed in Faraday cages when entering his companies' premises, effectively blocking all wireless signals to and from the devices. He then proceeded to bash Apple for its lack of “good models”, only to be fact-checked by his own platform.

This all happens against the backdrop of Microsoft having to recall its Recall feature , which was designed to operate on-device and serve a function similar to Apple Intelligence.

However, Apple’s move also serves as a prime example of how power consolidation can impact startups and scale-ups. The introduction of native AI features on Apple devices threatens to render many third-party applications, such as Grammarly and Goodnotes, redundant, as their core functions will be seamlessly integrated into pre-installed Apple apps.

Recognising the potential antitrust implications of this move, Apple preemptively announced that Apple Intelligence will not be available in the EU this year . This decision stems from concerns about potential breaches of the Digital Markets Act, which was designed to prevent monopolistic behaviour.

This trend of consolidation is also reflected in the recent 76%-drop of VC funding in Generative AI. Additionally, the company behind the Humane AI Pin, which faced multiple negative reviews , is now reportedly on the market with a $1 billion price tag .

Beyond privacy and antitrust concerns, the move towards on-device AI showcases a likely future of AI-human interactions: one where AI operates in the background, enhancing task performance and efficiency. Both Microsoft and OpenAI are actively vying for a similar role with their respective offerings Copilot and the new app OpenAI announced a month ago, but hasn’t released yet.

See the secretive AI device Johny Ive and Sam Altman are reportedly trying to create.

Apple's approach differs slightly, aiming to deeply integrate AI capabilities into all of its apps. Based on their recent demonstrations, Apple seems to be positioning AI both as an additional tool in its apps and as a traditional assistant (i.e., Siri).

See the paper Apple published on UI patterns for Apple Intelligence

As this landscape continues to evolve, it will be crucial to monitor how these strategies impact user experience, market competition, and how companies try to identify their unique value proposition. The balance among these will likely remain a central challenge in the AI industry.

Brain byte ??

We can break down an LLM-powered autonomous agent system in 5 modules:?an agent?with predefined instructions, a?memory?system for managing information storage and retrieval, specialised?tools?like web search, a?plan?module that develops strategies, decomposes objectives, and reflects on them, and an?action?module to execute decisions and plans.?

Although, not mandatory, it’s wise to introduce a?human-in-the-loop mechanism, too. That way you can ensure the agent behaves as intended at all times.

Want to learn more??Download our primer on AI agents.

Spotlight ??

Three main principles for designing conversational AI

The narrative around conversational AI taking over the digital world has been gaining momentum, especially after the release of GPT-4o and Project Astra.

Whether you believe this or not, one thing is certain—machine capabilities are finally catching up to the long-held human dream of the ever-capable AI assistant. And they’re about to change the way people expect to interact with your business.

The key to good conversational AI lies in designing conversations that go beyond a single medium. Think rich dialogues that guide users to make informed decisions and complete tasks effectively.

Here are three principles for designing useful conversational AI.

Want to learn more??Download our guide on conversational AI.

Sandbox ????

Each month we hand-pick a selection of models, tools and repos that can give you a hands-on experience with the latest developments in the field.

This month, we’re focusing on:

Tutorials

• Train GPT-2 in 90 minutes for $20
• Mistral’s cookbook full of tutorials such as building a RAG and fine-tuning a model

AI agents

• Anthropic’s agent tools Orchestrating sub-agents

Datasets

• HuggingFace’s FineWeb dataset with more than?15T tokens?of cleaned and deduplicated English web data from CommonCrawl
• Apple’s new image-text datasets

Video generation tools/repos

• Krea (freemium)
• Mora (open-source)
• Runway ’s new Gen 3 Alpha model (not freely available yet)

Modalities

• Google’s Deep Mind publishes new research on generating soundtracks for videos

Robots

• DORA (Dataflow-Oriented Robotic Application) is middleware designed to streamline and simplify the creation of AI-based robotic applications.

Our Watchlist ??

Some of the most interesting trends in AI this month:

Approaches to model explainability

• Anthropic’s research
• OpenAI

Tiktok

• The DAN (i.e., Do Anything Now) boyfriend trend on TikTok .
• Tiktok introduces a creative AI suite called Symphony

Robots and LLMs

• China’s ultra-realistic humanoid robots
• South Korea’s Starbucks robots
• Swallowable robots enter clinical trial

Geo-politics

• China’s UN resolution

The AI telenovela ??

Some call it the AI reality show, others Game of Thrones.

We prefer to call it “the AI telenovela” as no other term properly reflects the dramatic flair, back-stabbing, and absurdism that defines almost everything that’s been going on around AI monetisation efforts.

Enter our main protagonist (or antagonist - depending on your bend) - OpenAI.

Here’s what happened to OpenAI since the announcement of GPT-4o 1 month ago:

1. Sam Altman tweeted the phrase “her” insinuating that the voice Sky is similar to the one Scarlett Johansson played in the movie with the same name.
2. Scarlett Johansson released a public statement saying OpenAI reached out to her and she rejected their offer twice. OpenAI swiftly took the voice down
3. Altman kept making rounds on podcasts and conferences claiming that this was not intentional.
4. The voice actress behind ChatGPT's 'Sky' voice has come forward,?saying ?she was hired months before Altman contacted Scarlett Johansson.
5. A bunch of OpenAI’s security team left, including Ilya Sutskever (OpenAI’s co-founder who, reportedly, instigated the temporary firing of Sam Altman last year). One of the safety co-leads later on joined OpenAI’s rival Anthropic .
6. The team was dismantled, and now Altman leads OpenAI’s security efforts .
7. Then one of the ousted board members, Helen Toner, gave an interview saying that Altman was fired because he engaged in manipulative behaviour. She said the board learned about ChatGPT from Twitter.
8. OpenAI’s board said they’re disappointed Toner keeps rehashing the events.
9. Helen Toner and Tasha McCauley?penned ?an op-ed for The Economist arguing that AI companies cannot govern themselves.
10. In the meantime, reports came out that “manipulative behaviour” was a common thread in Altman’s career . His management team in the company he founded, asked the board to fire him for the same allegations. He was also fired from his Y Combinator position because he was not putting the interests of the company first.
11. Then, a report came out that OpenAI former employees had to sign aggressive non-disclosure agreements .
12. Altman said he was not aware of that and the issue will be fixed, but reports came out that both he and OpenAI lawyers knew about the NDA clauses.
13. Reports came out that although Altman has no financial stake in OpenAI, he was the head of the OpenAI fund that invested in promising startups.
14. Another report came out that a big chunk of the companies in which Altman’s major investor have already signed lucrative deals with OpenAI. For example, Reddit. OpenAI claimed that Altman has not been involved in any of these deals personally.
15. The co-founder who left OpenAI, Ilya Sutskever, has now founded a new company called Secure SuperIntelligence Inc (SSI).
16. Former NSA head joined the OpenAI’s board.

Meanwhile,

• Musk and Yann LeCun “debated” on X what science is while exchanging “pleasantries ”.

• Elon Musk dropped the lawsuit against OpenAI
• His xAI company raised $6 billion.
• Tesla shareholders filed a lawsuit against Elon Musk for breaching his fiduciary duty to Tesla by founding xAI.
• Meanwhile, the same shareholders reinstated Musk's $55.8B pay package and approved relocating the company to Texas.
• Meta’s position on open-source and its release of Llama made Zuckerberg popular again (he was even briefly considered an “Internet hunk” after a deepfake image of him circulated online ). In the meantime, Meta started developing its assistant, which will be trained on its users’ data. Only EU users have the right to opt out - a protection we get to enjoy thanks to GDPR.
• Finally, AI is getting firmer footing in military conflicts - see the former Google's CEO secret project of suicide drones that are already being tested in the US and Ukraine.

Stay tuned for the next episode!

Thanks for reading & see you next month!???

Iris Insights is written by the people at DEUS Humanity-Centered AI, a technology & design company focused on applied AI. Want to receive this newsletter in your inbox? Sign up here on LinkedIn or at iris.deus.ai .