After a bit of a crazy travel week, just four links, but here they are:
- Writer.AIs $100M Series B (the TechCrunch take is here): there have been a lot of AI mega-funding rounds, but I’m calling this one out because I think it’s a pattern we’ll see more of. A writing tool doesn’t seem eye-catching: how will they hold on to their audience, for example, if they are embedded in other UIs like editing tools? However, what Writer does is wrap their own custom LLM in tools that facilitate in-enterprise deployment (on-premises + SOC II compliant hosted) that can also tie in company-specific knowledge (databases, knowledge graphs, etc.). The intent is to be the primary writing productivity tool in large organizations, many of which are unlikely to want their employees making API Calls to OpenAI, Google, and others. It’s an enterprise sales play they’ll have to work through, but if they can establish themselves as a trusted AI within an organization, no doubt they can ask the long-term question: why stop at writing?
- OpenAI launches a Red Teaming Network: the company is looking for experts in a wide range of fields to help identify risk and safety issues in ChatGPT DALL-E. I wonder if the effort wouldn’t be better spent creating prompts, probes, and benchmarks that could be applied to all models (or at least all certain types). There are benchmarking efforts, but they are still small and barely funded.
- Anti Hype LLM reading list: a solid list of foundational reading to cut through the hype. Also, check the comments for a few more good pieces. There is so much content in AI it’s refreshing to have a curated list that isn’t part content marketing.
- Simulating history with hallucinations required: a nice reflective essay on experiments in teaching undergraduate history classes by simulating life situations from the distant past. Hallucinations can cause problems, but they also add color and ramp up student engagement. Part of the exercises created in the trials is for students to correct the hallucinations they see (flag them and write about them). That might be the biggest flaw: it seems enriching to have a historified story woven for learning, but it may be very hard to tell if subtle historical details are off.
Also, a fun update on the Facebook “don’t train with my data” process I mentioned here a few weeks ago. This turns out to be entirely useless (surprise, surprise). I received a response to my request to block this activity with a request for proof that the activity was happening. I selected the issue “I want to access, download, or correct any personal information from third parties used for generative AI.”, so I’d rather expect Facebook to provide me with who was using that data… (good luck with that apparently).
Have a great weekend!
