Aravind Srinivas

Want to share the thinking behind some of the recent UI changes on Perplexity that people may not have full clarity about:

Why Auto mode? - All AI products right now are shipping non-stop and adding a ton of buttons and dropdown menus and clutter. Including us. This is not sustainable. The user shouldn't have to learn so much to use a product. That's the motivation with "Auto" mode. Let the AI decide for the user if it's a quick-fast-answer query, or a slightly-slower-multi-step pro-search query, or slow-reasoning-mode query, or a really slow deep research query. The long-term future is that. An AI that decides the amount of compute to apply to a question, and maybe clarify with the user, when not super sure. Our goal isn't to save money and scam you in any way. It's genuinely to build a better product with less clutter and simple selector for customization options for the technically adept and well-informed users.. This is the right long-term convergence point.

Why are the models inconsistent across modes and why don't I see a model selector on Settings as before? Not all models apply to every mode. Eg: o3-mini and DeepSeek R1 don't make sense in the context of Pro Search. They are meant to reason and go through chain-of-thought and summarize; while models like Sonnet-3.7 (no thinking mode) or GPT-4o are meant to be really great summarizers with quick-fast-reasoning capabilities (and hence good for Pro searches). If we had the model selector in the same way as before, this just leads to more confusion as to which model to pick for what mode. As for Deep Research, it's a combination of multiple models that all work together right now: 4o, Sonnet, R1, Sonar. There's absolutely nothing to control there, and hence, why no model choice offered.

How does the new model selector work? Auto doesn't need you to pick anything. Pro is customizable. Pro will persist across follow-ups. Reasoning does not, but we intend to merge Pro and Reasoning into one single mode, where if you pick R1/o3-mini, chain-of-thought will automatically apply. Deep Research will remain its own separate thing. The purpose of Auto is to route your query to the best model for the given task. It’s far from perfect today but our aim is to make it so good that you don’t have to keep up with the latest 4o, 3.7, r1, etc.

Infra Challenges: We're working on a new more powerful deep research agent that thinks for 30 mins or more, and will be the best research agent out there. This includes building some of the tool use and interactive and code-execution capabilities that some recent prototypes like Manus have shown. We need a rewrite of our infrastructure to do this at scale. This meant transitioning the way we do our logging and lookups, and removing code written Python and rewriting it in GoLang. This is causing us some challenges we didn't foresee on the core product. You the user shouldn't ideally even need to worry about all this. Our fault. We are going to deprioritize shipping new features at the pace we normally do and just invest into a stable infrastructure that will maximize long-term velocity over short-term quick ships.

Why does Deep Research and Reasoning go back to Auto for follow-ups? - Few months ago, we asked ourselves “What stops users from asking follow-up questions?” Given we can’t ask each of you individually, we looked at the data and saw that 15-20% of Deep Research queries are not seen at all bc they take too long; many users ask simple follow-ups. As a result, this was our attempt at making follow-ups fast and convenient. We realize many of you want continued Reasoning mode for your work, so we’re planning to make those models sticky. To do this, we’ll combine the Pro + Reasoning models as “Pro”, which will be sticky and not default to Auto.

Why no GPT-4.5? - This is an easier one. The decoding speed for GPT-4.5 is only 11 tokens/sec (for comparison, 4o does 110 tokens/sec (10x faster) and our own Sonar model does 1200 tokens/sec (100x faster)). This led to a subpar experience for our users who expect fast, accurate answers. Until we can achieve speeds similar to what users expect, we will have to hold off on providing access to this model.

Why are there so many UI bugs & things missing/reappearing? - We’re always working to improve the answer experience with redesigns, like the new Answer Modes. In the spirit of shipping so much code and launching quickly, we’ve missed the mark on quality, leading to various bugs and confusion for users. We’re unapologetic in trying new things for our users, but do apologize for the recent dip in quality and lack of transparency (more on that below). We’re implementing stronger processes to improve our quality going forward.

Are we running out of funding and facing market pressure to IPO? No. We have all the funding we've raised, and our revenue is only growing. The objective behind Auto mode is to make the product better, not to save costs. If anything, I have learned it's better to communicate more transparently to avoid the any incorrect conclusions. Re IPO: We have no plans of IPOing before 2028.

Lastly, to provide more transparency and insight into what we’re working on, we'll be doing monthly AMAs for users here to answer your questions. And share changelogs more frequently.