Great, but the elephant in the room is you can't trust these things to tell the truth, and given that, I struggle to see how you can use these for anything other than creating fancy emails.
Databricks claims its open source foundational LLM outsmarts GPT-3.5
Analytics platform Databricks has launched an open source foundational large language model, hoping enterprises will opt to use its tools to jump on the LLM bandwagon. The biz, founded around Apache Spark, published a slew of benchmarks claiming its general-purpose LLM – dubbed DBRX – beat open source rivals on language …
COMMENTS
-
-
-
Tuesday 2nd April 2024 14:17 GMT Ken G
You're thinking small, how about an individualised chatbot that messages you (or calls, through that'd be more expensive to set up) with reasons why you should vote for their local candidate based on where you went to school, where you work, what age you are and what family you have etc. maybe pretending to be an old friend you haven't heard from in years.
-
-
Friday 29th March 2024 15:42 GMT Michael Wojcik
There are some legitimate applications in computational science and mathematics, along the lines of "show me something surprising and I'll see if it checks out" (but of course much more complicated). The work that's being done with elliptic-curve murmurations is an example. Autoregressive transformer stacks have also been used to suggest chemical formulations that may have properties of interest, for example.
But yes, in general, using a tool that is still far beyond the frontier of explicability is a Bad Idea in most domains. And explicability isn't improving all that quickly. Sparse auto-encoders and variational-quantized auto-encoders can provide some information; so can linear probes. There's been some interesting discussion lately on using AtP* and (separately) Tracr for interpretation and explication. We're still very far from being able to make even weak explanatory claims about the behavior of SotA models, though.
-
Saturday 30th March 2024 10:01 GMT Mike007
This week I was given the task of evaluating PrivateGPT as a potential product offering. I tried feeding in the product specification documents for a manufacturers entire product range. I gave it a scenario and asked for a product recommendation.
At least it was able to tell me which page of which document contained the specifications for the 20mm bolt it was recommending for a 16mm hole...
-