I have some serious issues with this company's PR.
They say their Inflection-2.5 model is the world's best personal AI[0] which is a dumb claim to make considering it is done off of automated benchmarks which we know are flawed and even if you assume automated benchmarks are good enough to claim that title, it would be held by dozens of other open weight models on HF, not Inflection-2.5.
They said their Inflection-2 model was the second best in the world[1] while comparing it to Palm-2 which no one considers close to the best in the world. They again, based their claim on automated benchmarks which anyone knowledgeable in the space would know can be gamed and is not representative of actual conversational performance. (take a look at the Lmsys Arena Leaderboard for a better metric)
They list other models they consider good while failing to mention or compare to Mixtral 8x7b, the best open model that exists.
And they introduce buzzwords no one in the area uses like IQ and EQ as if they are innovative concepts.
Making big, bold claims without evidence is the exact kind of manipulative PR speak I'd expect from a company with little to no substance.
"After just 3 days of being incorporated and just 2 days of actual work performed, we have now created the best AI on the planet with trillions of active subscribers, per second.
The quality of our product is so high, it is being compared to a perfect diamond the size of the entire solar system. But, we are humble, and while we think our tech is pretty darn good, we know that we can do better. That's why we are introducing v2 in just 2 business days from today, which will render all other forms of intelligence (human and artificial) irrelevant."
It really captures what I felt as I started to read this PR bullshit. I fed it through an LLM to summarize it, and there was no substantive content in that summary.
Lampooning aside, the sobering reality is that a tiny number of people are acquiring O($10^8) wealth (or more) via such shenanigans, and that is the reason we will see a lot more of this.
Dunno, it's pretty damn good and quite useful. I asked Mistral 7b and Claude Opus to count the banned substances on the WADA anti doping list. Both avoided a straight response giving back loads of bs. Claude was next to useless. Mistral gave a somewhat acceptable answer counting substances on every chapter after much persuasion. Inflection's pi AI gave me a straight answer to the first question without any bs. I like its casual tone and the fact that it does not use needlessly superfluous language like the GPT models.
Also, and I'm going to put this real bluntly, a "personal AI" isn't worth crap to a lot of people until it's allowed to talk about sex.
I'm not even talking for porn purposes. Sex is a basic and healthy part of most peoples' lives, any number of relationship issues revolve around it, and for plenty of people it's literally their profession, legal or not. Anything that pretends it can 'personally' help a broad spectrum of people while treating sex as verboten is just bullshit.
My firm belief is that there's no moral objection here by the companies involved. They just want to avoid being cut off by credit card processors because porn, and are cowardly enough to act like doing so is somehow protecting users.
They learned that most people can't tell the difference between actual technical concepts and PR speak and it worked to sell NFTs, so now that AI is all the hype, guess what.
They say their Inflection-2.5 model is the world's best personal AI[0] which is a dumb claim to make considering it is done off of automated benchmarks which we know are flawed and even if you assume automated benchmarks are good enough to claim that title, it would be held by dozens of other open weight models on HF, not Inflection-2.5.
They said their Inflection-2 model was the second best in the world[1] while comparing it to Palm-2 which no one considers close to the best in the world. They again, based their claim on automated benchmarks which anyone knowledgeable in the space would know can be gamed and is not representative of actual conversational performance. (take a look at the Lmsys Arena Leaderboard for a better metric)
They list other models they consider good while failing to mention or compare to Mixtral 8x7b, the best open model that exists.
And they introduce buzzwords no one in the area uses like IQ and EQ as if they are innovative concepts.
Making big, bold claims without evidence is the exact kind of manipulative PR speak I'd expect from a company with little to no substance.
[0] https://inflection.ai/inflection-2-5
[1] https://inflection.ai/inflection-2