Sunday Links: High Risk AI, Small Language Models, and AI Certifications

Welcome to another round-up of interesting AI news and some general web news as well. Here they are:

• Google Gemini dubbed ‘high risk’ for kids and teens in new safety assessment. Together with OpenAI being sued for user deaths within, this is another story on how LLM controls for safety really don't work that well yet. Most of the controls and guardrails are added in post-training or within guardrails during usage. These mechanisms are just too weak to work yet. There are hypotheses that putting stronger ethical guidelines into the LLM earlier in the process will be more effective, but there is clearly a lot of work to be done.
• How Small Language Models Are Key to Scalable Agentic AI. NVIDIA hypes up Small Language Models. I happen to agree with NVIDIA that small, efficient language models for specific tasks are likely to be much more effective for many tasks than large general models. What's not clear to me is why NVIDIA is promoting this as well? One would imagine they benefit from training and inference with large models. Perhaps, though, the game is that with smaller models, the number of organizations that can train and run their own models is so much greater that NVIDIA can grow faster and have a more diversified customer base.
• Exa raises $85M to build the search engine for AIs. Exa is one of a new generation of 'tool' services for AI. I've written about them before (see "Sunday Links: AI marriages, voice, and tools for AI agents"), and the company focused on semantic indexing of content. The idea is that semantic searches are much easier to drive from an LLM agent (and serve an LLM agent better) than the Google 10 blue links search. The answer as to why is clear: when faced with 10 blue links, an LLM is still faced with the next step: click them all and aggregate the content. It turns out, though, that the same is true for humans. I've used Exa extensively to research products and companies. It's not perfect, but getting a much deeper analysis of each item in the list that's being returned saves hours of work.
• AI company Anthropic agrees to pay $1.5B to settle lawsuit with authors. This news has been widely covered, and the reaction ranges from $1.5B being a record settlement to this resolving little, since the settlement relates to the pirated use of material Anthropic did not own, rather than the legality of training on copyrighted books. The latter has not been prohibited by the courts in the United States, and it also appears that Anthropic will not be required to "detrain" or delete its trained models (which would have had huge implications). Given those facts, Anthropic may be quite happy with the outcome (and given that the company just raised $13B, payment should not be a problem).
• Expanding economic opportunity with AI. In probably the most perplexing news of the week, OpenAI announced an AI jobs platform and an AI certifications program. The aim of the new service is to match AI talent with businesses that need new skill sets to survive the coming AI economic disruption. It's hard to tell if this is a new revenue stream, a way to tie the idea of "AI competence" to ChatGPT use, or a political nod towards upcoming economic impact. Neither of the services is live yet, but it seems to me that there will be little mileage in the idea of "AI certifications". Applications of AI and tooling will be so diverse that this will likely be meaningless. With AI technology also moving so fast, it's hard to see how this will be grounded in any kind of stable footing. It seems a little like a certification in Internet usage.

Finally, the biggest tech "non-AI" news this week was likely the release of the ruling in the recent US Google antitrust case. Judge Amit Mehta had already ruled that Google had abused its monopoly position by paying Apple and others for exclusive rights to being the default search engine on devices. The remedies could have been severe (though in truth, forcing Google to divest the Chrome browser or Android always seemed unlikely). Most observers feel that the remedies as stated are quite light. In essence ,the judge ruled that Google could continue to pay for placement of search (though not exclusively) in large part because the web ecosystem depends on these payments. There is a part of the ruling that Google must share its search index, however, at least twice, but perhaps more often. It's hard to figure out what that will mean in the business of search, but perhaps the biggest winners in this case will be OpenAI and Meta.

Wishing you a great weekend.