Personal AI got more powerful, courtesy of Inflection 2.5

Inflection AI, the AI company by DeepMind's Mustafa Suleyman and LinkedIn's Reid Hoffman, has marked its attendance into the “GPT-4 class” of models. Their new model, Inflection-2.5, goes toe-to-toe with giants like GPT-4, all while saving a ton on training costs.

What’s going on here?

Inflection AI’s new model, Inflection 2.5, competes with GPT-4 to build the ultimate personal AI.

What does that mean?

Honestly, people might have forgotten about Inflection AI for a minute. They promised a different kind of personal AI and talked about building the largest cluster of GPUs. Well, they've seriously delivered. This Inflection-2.5 model is a game-changer, and the numbers speak for themselves.

They're claiming it does 94% of what GPT-4 can do using way less training power (and let's be real, nobody knows exactly what OpenAI spent on GPT-4!). On top of raw smarts, it's crushing it in areas like coding, math, and general IQ.

Inflection’s free AI chatbot, Pi, has already got Inflection 2.5 running in the background. The stats for Pi’s usage are another shocker. It has 6M MAU with an average chat being over 33 minutes.

It’s also growing 10% month over month. And now, it also has access to real-time info from the web. Pi is on all platforms—desktop, iOS, and Android. Pi’s voice output makes it really feel like a friend (with that quirky personality—see image below).

ps: some replies from Pi are eerily similar to Anthropic’s Claude. Are they A/b testing with Anthropic’s API? That’d be hilarious.

Why should I care?

It’s surprising to get beastly models, actually 2, from AI companies (Anthropic and Inflection) that lean on the AI safety and regulation side. In an exclusive interview with Axios, Mustafa mentioned, "It's been an amazing story of access, I would argue, and actually a very impressive story of very limited harms.”

Apart from that, this release opens up many possibilities:

• You can train powerful models with less compute.

• Inflection also released a corrected version of MT-bench, which they found has 25% inaccurate answers.

• There’s demand for personal AI chatbots, not just work assistants.

Well, good for us, now maybe OpenAI will release whatever they have been cooking.