Reporting notes from the AI trenches, Jan 29 - Feb 2
From Taylor Swift deepfakes and open source to Meta, Salesforce and the news I didn't have time to tackle
Sometimes by the end of a week of reporting for VentureBeat, I’m amazed at how much AI news I managed to cover — and even more amazed at how much I didn’t have time for. This week was no different.
Monday/Tuesday:
I started out, as I often do, with a strong idea of what I wanted to tackle in my weekly AI Beat column, which I publish on VentureBeat and also include as part of a LinkedIn newsletter. I’ll typically think about it over the previous weekend and then publish on Monday or Tuesday.
I knew that I wanted to tackle the Taylor Swift deepfakes. And as someone who recently saw the Eras tour movie with my niece, and has recently done a deep dive into all of her albums, it was easy to come up with a serious take on the issue — that AI companies will not simply be able to ignore it — but there was no way I could help but add some song titles into the mix, like ‘Shake it Off,’ ‘Anti-Hero’ and ‘Bad Blood.’
Wednesday:
Last week I did a terrific interview with Clara Shih, CEO of Salesforce AI, and I was excited to put together an article about our chat, which I published on Wednesday.
Shih had a great story to share about her own personal generative AI ‘aha’ moment back in November 2021, at a meeting with an Italian delegation from fashion brand Gucci where Salesforce chief scientist (and previous Stanford professor) Silvio Savarese gave the group their first look at large language models.
But she also gave me a clear idea of that helped Salesforce to get its EinsteinGPT offering shipped just a few months after ChatGPT debuted in November 2022 — and how even as AI remains a ‘moving target,’ Shih keeps her eye on the prize.
Thursday:
By Thursday, my email inbox was bursting, as usual — but I got one last reminder from the folks at the Allen Institute for AI (AI2) about their new OLMo LLM, a ‘truly’ open source model — it doesn’t just include the model code and model weights, but also provides the training code, training data and associated toolkits, as well as evaluation toolkits.
Once I published my story, however, it turns out I had made an error — the press kit had made it sound like the model was the ‘first’ of its kind, but in hindsight I misunderstood and should have looked at that more closely. The fact is, OLMo builds on the work of other fully open source models like EleutherAI’s Pythia and Big Science’s BLOOM. It is only the ‘first’ OLMo LLM.
Later that evening, EleutherAI executive director Stella Biderman called me out on that — and rightly so — and I quickly made a correction to the story. The good news is, I ended up getting a fun fact about OLMo that I didn’t know. AI2 researcher Luca Soldani let me know on Twitter/X that “the name of the pretraining corpus, Dolma, stands for Data to feed OLMo’s Appetite.” Oh, AI researcher humor!
Friday:
I worked on one last quick piece on Friday, after I saw Alex Heath from The Verge post about an “insane stat” from Meta’s earnings call, in which Mark Zuckerberg estimated that the company’s public user data available for AI training is greater than the internet's Common Crawl dataset, '“which is over 6 PETABYTES.”
I went through the Meta earnings call transcript and found it fascinating start to finish — with the deepest dive of all the Big Tech companies into its unique AI strategy. Zuck clarified Meta’s play-to-win strategy in AI, which includes a strong focus on building a ‘world-class compute infrastructure’ for AI; an open source/open science AI strategy that is front and center; and, of course, that juicy, valuable, vast store of training data in the form of Facebook and Instagram public posts and comments.
What I missed:
There was so much I couldn’t get to this week — but I wanted to! So I’ll share some links to the most important stories:
The EU AI Act — it’s a done deal, at least until it is implemented (which will take a while)
The White House’s update on actions following Biden’s executive order on AI — I really wanted to get to this, but alas. To be honest, I haven’t even had a chance to read it. That’s how impossible this beat is sometimes, lol.
Hugging Face’s new GPT maker — my VentureBeat colleague Carl Franzen tackled this one. A free competitor to OpenAI’s custom GPT builder? Hey, can’t argue with free, and you can choose your own open source model to use.
Amazon’s new AI shopping assistant, Rufus — if Amazon’s shopping experience weren’t so hard to plow through these days, maybe this wouldn’t be necessary. But I’m willing to give Rufus a walk (it does sound like a dog name, doesn’t it?)
I’m got one more day of weekend to enjoy, but I’m already thinking about my next AI Beat column — keep an eye out for it on VentureBeat Monday or Tuesday.
Sharon :)