Recall that circa 2015-2016 AI was about to replace half of all jobs, including all drivers, most doctors, etc. People's perception of AI progress is rarely grounded in actual capabilities — people are always projecting their hopes on the latest hype trend (e.g. deep RL)
@fchollet
-
Deep Learning’s Economic Impact Falls Short of Expectations
By
–
There's a parallel to deep learning itself — the economic impact of the tech has been ~10% of what folks expected ~8 years ago, and being a provider of deep learning APIs / models has been a rather lousy business. Although individual engineers have done very well for themselves
-
Deep Learning Models as Knowledge Retrievers and Action Routers
By
–
Yes, once DL models are used not as raw knowledge stores (which doesn't make a lot of sense) but as knowledge retrievers and action routers, they will be as up-to-date as the underlying database. https://
t.co/yHN13PyJeA -
Tech Commoditization and Engineer Value in Deep Learning
By
–
The more useful the tech turns out to be the more it will get commoditized and the harder it will be to monetize (paradoxically). The only clear winners are individual engineers with great deep learning NLP skills.
-
AI Impact Will Be 10% of Inflated Public Expectations
By
–
With all that said… it's obvious to me that the actual impact of the tech will be maybe ~10% of what the average person on my timeline expects. People have *ridiculously* inflated expectations, that aren't grounded in the actual capabilities (current or future) of these models.
-
Future Models: Local Conversational Interface with Server-Side Backend
By
–
I wouldn't be surprised if we end up with models that run *locally* (in the browser or on your phone), acting as a conversational interface between you and a server-side knowledge/task backend.
-
Future AI assistants will be more capable and cheaper to operate
By
–
That means that the AI assistants of the near future will be considerably more capable than what we see today. And they will be a lot cheaper to run too, as model distillation techniques and architecture refinements keep catching up to raw model size.
-
Integrating LLMs with Symbolic Tools to Address Limitations
By
–
Likewise we can interface LLMs with an array of symbolic tools that can shore up their weaknesses — calculators, interpreters, discrete search programs, SAT solvers, etc.
-
LLM Performance Evolution: Beyond Generation to Knowledge Retrieval
By
–
It's clear that we're far from peak LLM performance — these models will keep getting better. It's also clear that pure generation is just the first step — we can largely alleviate the LLM reliability issue by using them as information retrieval devices over a knowledge corpus.
-
KerasNLP APIs for Custom Tokenizers and Advanced NLP Use Cases
By
–
At an even lower level, you can create your own custom tokenizer on a new vocabulary, or pretrain your own backbone. The KerasNLP APIs will be useful to you no matter how advanced you use case becomes. Be sure to check out the starter guide: https://
keras.io/guides/keras_n
lp/getting_started/
…