For the past several months developers have been obsessing over performance scores whenever a new LLM comes out. That may be changing, though, as some developers increasingly look for alternate ways to evaluate the success of products powered by AI.
@mattlynley
-
Open Source AI Models: Predictability as Key Advantage
By
–
But it's adding a new layer to the discussion around the benefits of open source models. In addition to the usual stuff—latency, cost, privacy—they're increasingly attractive because they're just predictable.
-
API Dependencies: Trading Control for Convenience and Speed
By
–
That's, of course, the bargain you make when building a product on top of any API: you trade convenience and agility for control of the technology under your products. But if history is any indicator, a whole universe of companies are willing to make that trade.
-
OpenAI Faces Hardware Constraints Limiting GPT-4 Access
By
–
OpenAI is, like many other companies, facing hardware constraints. It's directionally honest about it. It limits the number of GPT-4 calls for Chat users. Developers have to pay for its APIs (like davinci or GPT 3.5-Turbo) for at least a month before they get access to GPT-4.
-
GPT-4 API Reliability Issues and Rate-Limiting Challenges
By
–
One theme that comes up increasingly with sources and experts I speak with lately is a complicated challenge for OpenAI: Reliability. Or, more specifically, that people trying to use the GPT-4 API are increasingly running into performance issues and rate-limiting.
-
Manual Data Processing Humorously Called Machine Learning
By
–
Me: *methodically clicking around and copy/pasting for three hours into a spreadsheet* Them: whatcha doin? Me: machine learning
-
Meta’s Research-Based Recruiting Strategy Revealed
By
–
Supervised readers would have caught wind of this in June 🙂 https://
supervised.news/p/research-as-
recruiting-and-metas
… -
Coining ‘Zirpy’ to Describe the Bonkers Zero Interest Rate Era
By
–
Petition to make zirpy an adjective to describe in retrospect how fucking bonkers the zirp era was
-
RAG and Agents Drive Llama 2 Potential Forward
By
–
And then we've also seen both the rise of RAG and increasing interest in agents—both of which feed directly back into the potential of Llama 2. Of course there's plenty more that happened, but what I like to cover is not just the news—it's what people are talking about.
-
OpenAI Extends Deprecation Window Amid Developer Backlash
By
–
We also saw OpenAI capitulating to developers by extending a deprecation window to a full year (up from 3mo) due to "feedback" (aka backlash) from developers. It also saw the rise of a narrative of GPT-4's performance dropping, which was a bit of a misinterpretation of a paper.