Weekly Link Roundup #30

Has it been 30 already? Gosh, that's the Zinc on the Periodic Table!

A word about these two images…all I did was put “30” into both Dall-E (above) and Midjourney (below). Dall-E’s output I get. MJ’s output is easy to dismiss with a joke about “are you OK MidJourney? You seem a little freaked out”..but the output just makes me wonder how the input is being translated into the ask from the training data. I find it hard to believe that those images correspond to some predictive outcome from “30.”

So then I asked MJ for “the number “30” and got….

I mean I had “30” in quotes and it added an extra zero to one. So I’m still puzzled by how these two are interpreting the input but for now, on with the newsletter.

Capgemini digs into the real reasons that gen AI proof of concepts rarely take off: I really like this article for just this bit, “He went on to explain that a big chunk of the reason that data is often referred to as the new oil is because oil’s only useful after refinement.” Could not be more spot on. Couple of points I’ve routinely hit - 1) people should be experimenting with AI in their orgs - that means time and resource bound with criteria for success 2) Senior Leadership need be more technically savvy about an AI stack than with any other technology 3) once you get past the pilot stage, you run smack into the data wall and a lot of orgs are going to find out that they’re sitting on a pile of messy, bad data that will not help at scale. Data cleanup operations need to be in everyone’s budget right now.

How the Stream Deck rose from the ashes of a legendary keyboard: Another story to love. What’s not to love? The story of how compelling concept images can drive downstream breakthroughs - if you can see it, you can build it. A story of how trying to innovate on something as seemingly dead-end as a keyboard can lead to the creation of essentially a whole new market sector with nothing to do with keyboards. It’s tempting to get so heads down that we forget to look around and not just around but outside of our silos.

Cohere teams up with Fujitsu to launch Japanese LLM ‘Takane’ for enterprises: Strong signal for two reasons - 1) while the stock price of AI-enabling companies might be in a bubble, the tech is not and it will continue to spread and give birth to more and more niche and not so niche foundational models and 2) one area that will continue to grow will be in foundational models built in native languages other than English.

Mistral releases Codestral Mamba for faster, longer code generation: This is a signal for me but for what might be a non-obvious reason - it’s another indicator of how smart leaders will need to be to judge the value of implementing LLMs for a range of activities.

Microsoft’s new AI system ‘SpreadsheetLLM’ unlocks insights from spreadsheets, boosting enterprise productivity: Who remembers VisiCalc? if you don’t, look it up. That spreadsheet program was one of the main propellants of the move to PCs. This shares a notable characteristic with its ancestor - people get it - they get doing stuff with data in spreadsheets in a way that carries less of a cognitive load than in other applications. Another signal boost here is that SOOOOOOOO much corporate data is lodged within spreadsheets, that a purpose-built LLM to work with that data has some serious potential. Add in the fact that its from MSFT which means its already inside most enterprises and this is a sleep hit waiting to happen.

Vectara raises $25M as it launches Mockingbird LLM for enterprise RAG applications: Look, I know I talk A LOT about AI but there is a lot to talk about. RAG (retrieval augmented generation) -basically the capability of an LLM to access data outside of its training set is a key to moving the utility of LLMs forward, especially in the enterprise. Think about an LLM tasked with providing learning content to a sales force. When RAG is implemented, you could point the LLM at data coming in fresh from the field vs just working off the training data set. Your content just got fresher.

How Europe’s universities are using AI to battle dementia: Wow > > “The scientists validated the AI tool’s predictions with follow-up data over the course of six years. They suggest that their solution is three times more accurate at predicting Alzheimer’s progression than clinical diagnosis or clinical markers such as grey matter atrophy and cognitive scores.” Early detection is always key and if this kind of tech advancement can battle this wretched disease at all, I will cheer for this team as hard as possible.

Intel launches apprenticeship for U.S. manufacturing workers: Good for Intel. In the AI-powered world of needing to reskill and upskill so much of the workforce, apprentice programs like this could be a really valuable approach - “Selected apprentices will be full-time Intel employees on day one and will earn a certificate and college credit upon successful completion of the one-year program.” > > Now I’m thinking about a Guild model where the Guild is almost like a union but for holding professional qualifications outside of an employer - what if you amass validated skills through apprentice programs instead of degrees?

What The Jetsons got right and wrong about the future of work: Most amazing part of the whole story? The Jetsons were only on for 1 season.

The Prompt Report: A Systematic Survey of Prompting Techniques: Seems helpful > > “This paper establishes a structured understanding of prompts, by assembling a taxonomy of prompting techniques and analyzing their use. We present a comprehensive vocabulary of 33 vocabulary terms, a taxonomy of 58 text-only prompting techniques, and 40 techniques for other modalities.”

INTERNET OF AGENTS: WEAVING A WEB OF HETEROGENEOUS AGENTS FOR COLLABORATIVE INTELLIGENCE: (PDF link) “Inspired by the concept of the Internet, we propose the Internet of Agents (IoA), a novel framework that addresses these limitations by providing a flexible and scalable platform for LLM-based multi-agent collaboration.”

This HR company tried to treat AI bots like people — it didn’t go over well: No one saw that ending badly. /smh

An Education Chatbot Company Collapsed. Where Did the Student Data Go?: In the sense that failures can be instructive, this is a whopper. It includes lessons on data privacy, how good tech doesn’t mean a good business, how hard it is to scale, and how leaders have never needed to be smarter about the tech they’re deploying and its implications, than ever before.

How technology is reinventing education (Stanford Report): Going to be a little petty here and include this just so I can disagree with it. Just like NO ONE is going to “reinvent HR” unless we ditch 800 year-old double entry bookkeeping that ONLY allows for employees, there will be NO reinvention of public education unless we change the way it is funded. Period.

‘Death Occurs in the Dark’: Indie Video Game Devs Are Struggling to Survive: This makes me sad thinking about all the creative work that is being lost.

New Senate bill seeks to protect artists’ and journalists’ content from AI use: Hmmm…*rubs chin thoughtfully in government* “The bill would require companies that develop AI tools to allow users to attach content provenance information to their content within two years. Content provenance information refers to machine-readable information that documents the origin of digital content, such as photos and news articles. According to the bill, works with content provenance information could not be used to train AI models or generate AI content.”

Atropos Health lands $33M to scale AI-powered real-world evidence, build out pharma partnerships: We need an index of which industries have the “cleanest” data > > “Queries pull on the company’s network of hundreds of millions of de-identified patient records. The system then runs the analysis using statistical methods and then formats the final result into an observational study report. Such studies can traditionally take weeks or months, the company noted.”

Cloudflare launches a tool to combat AI bots: OK so this is a strong signal to me of the battle to defend content from becoming part of an LLM training set. It’s strong because Cloudfare operates at near-infrastructure levels and what they do can impact large chunks of the Web at a single pass. There’s always a catch though and in this case it’s the market power that Google has in the ad market “Tools like Cloudflare’s could help — but only if they prove to be accurate in detecting clandestine AI bots. And they won’t solve the more intractable problem of publishers risking sacrificing referral traffic from AI tools like Google’s AI Overviews, which exclude sites from inclusion if they block specific AI crawlers.”

This Photographer Documented Native American Tribes Before They ‘Vanished’: (really appreciate the use of quote marks around vanished - no, I’m not being sarcastic - these cultures have not vanished but this photographic record is still important) “The work of an early 20th-century photographer who set out on a 30-year mission to document Native American tribes was almost forgotten about entirely until his work was rediscovered years later.”