The AI Strategist - 12.5.23
We don’t need no AI-ducation
In the news - on 12th May 2023 edX, a leading global online platform, launched two new Generative AI innovations. A ChatGPT plugin and an AI-powered learning assistant. Powered by OpenAI, these products aim to guide the learner and assist with discovery. edX have promised many more future innovations that will ‘harness the power and potential of Generative AI.’
The first product takes advantage of ChatGPT plugins (which were launched on March 23rd 2023) to improve content discovery. The second product leverages the ChatGPT API to help the individual learn and address queries.
Why this matters - edX have moved quickly to capitalise on the potential of foundation models (LLMs). Chegg, another edtech provider, grew significantly during the pandemic. In 2021 Forbes declared them the most valuable edtech company in America. On May 2nd 2023 they lost close to £1bn in market cap based on the threat of Chat GPT to their business model. Shares in Pearson, a competitor, fell 12% on the same day.
Generative AI is disrupting business models at pace and organisations are looking to leverage foundation models quickly as a defensive and reactive strategy. We are seeing product innovation and launch cycles rapidly reduce (i.e. from 6 months down to 6 weeks). This puts those providing these models (i.e. OpenAI) in a very strong commercial position. As foundation models become commoditised, those organisations thinking strategically about their positioning in the Generative AI stack will win in the long run.
2. LegislAItion
In the news - on 11th May 2023 European Union lawmakers approved amendments to the EU AI Act. This means it now incorporates safety and transparency rules for Generative AI. The text was backed by an 84-7 vote, with 12 abstentions.
MEPs propose that Generative AI be held to higher standards of transparency and human rights compliance, and that developers partly disclose what copyright-protected content was used to train these models.
Why this matters - this is a significant development because the rest of the world are looking to Europe to set the precedent for AI regulation. The progression of the EU AI Act was slowed down following the launch of ChatGPT in November 2022. In the absence of AI regulation, organisations are self-policing through a combination of ethical principles (soft rules) and other relevant legislation such as GDPR (i.e. hard rules).
I can recommend this book if you want to read more on the topic.
Further reading
The Ada Lovelace Institute presents a good summary of the challenges Foundation Models (also referred to as General Purpose AI - GPAI) have introduced when it comes to AI Governance. Distribution routes:
Route 1 - i.e. AI organisation accesses GPAI via an API
Examples of GPAI models distributed via API are OpenAI’s GPT-3.5 (and its user-facing system, ChatGPT) and DALL-E. The key feature of this route is that control over the model and source code remains largely in the hands of the provider.
Route 2 - i.e. AI organisation accesses GPAI via open-source access
Open-source access refers instead to releasing the model or some elements of it publicly and allowing anyone to download, modify and distribute it, under the terms of permissive licences. In this case, only a one-off interaction between the GPAI provider and the downstream developer is needed. Stability AI and RunwayML adopted this release strategy with Stable Diffusion.
Challenge for regulators - regardless of route, any organisation building a Generative AI application is considered ‘downstream.’ These organisations have a responsibility to deliver safe and ethical AI applications when adapting the GPAI model. If they want to use training data for fine tuning or evaluating the model then they are more restricted in route 1 in comparison to route 2.
In terms of what the ‘optimal’ safety/ethics outcome is for society there are arguments to make for the merits of both routes. The real complexity stems from the fact that we are in the very early stages of GPAI monetisation. As a consequence, many more release strategies will emerge. The key challenge for regulation is to future-proof policy by making it both adaptable and independent of GPAI release strategies.
3. In the PaLM of your hand
In the news - on 10th May 2023 Google had their annual developer conference - Google I/O.
You can watch the keynote here.
The big news was about PaLM 2 (LLM). There’s 91 pages of further information here if you want more detail. The training dataset is web documents, books, code, mathematics, and conversational data. It is notable that this dataset ‘includes a higher percentage of non-English data than previous large language models.’
Why this matters - although some of the media commentary focused on the lack of transparency regarding the training data, citing a strategic shift based on Microsoft/OpenAI competition, I was delighted to see the last 30 pages of the 91 page document dedicated to Responsible AI.
As I mentioned above, there is a lot of reactive innovation happening in response to the ubiquity and price point of Chat GPT. Generative AI innovators clearly need to think strategically about the tech stack and how this will play out. They also need to be thoughtful and deliberate about their Safety & Ethics strategy. A big part of this is who you align to when it comes to foundation models and how you collaborate with them.
As an aside - CHAPTR, the company I have been working with since October 2022, were featured in the Google keynote. You can watch a short video of this here. We have been collaborating with Google from the start and will be using their technology in our first product - STORI.