Comparing LLMs, Bloomberg Odd Lots, and More.

Update Email

Jul 21, 2025

A few updates for y’all.

Iqidis was featured on the respected Odd Lots Podcast by Bloomberg (Spotify linked here). We continue to make waves w/o relying paid PR, paid promotions, or other “growth tactics” that our competitors need.
Future Plans for the newsletter— live streams and an interactive AI podcast. Suggest your timings
Sharing this post by the exceptional
Sebastian Raschka, PhD
with a thought about how this can change how we use intelligence in AI.

Iqidis on Bloomberg

Iqidis was spontaneously mentioned in Bloomberg’s investment/finance focused podcast, Odd Lots, when interviewing Joel Wertheimer, a notable civil rights lawyer. When asked about the Legal AI tools he actually liked, Joel said that Iqidis looked promising.

This continues our amazing word-of-mouth-fueled growth. While our competitors are burning millions of paid PR and forced outreach, we’ve attained word-of-mouth virality by obsessively focusing on the product and in delivering premium customer service to our customers. We’ve relied on 3 pillars to unlock with <200K what our competitors have failed to achieve w/ hundreds of Millions (use your favorite deep research tool and run “Iqidis vs Harvey User Feedback”, you’ll see the massive difference in user sentiment).

We’ve done this through 3 pillars-

Open Source: We leverage open source AI and Research communities aggressively to cut down our own RnD costs. While our team on paper might be small, we’re aggregating insights from a huge network of experts. This allows us to stay on (and in many cases redefine) the cutting edge of AI for Legal.
Premium Customer Support: We treat our users with the utmost respect, and they can feel that. We’ve been open in acknowledging fault, revert quickly to all queries, and add user-requested features at breakneck speed. These are all small things, but they add up over time. It’s why our users actively petition for us.
Accessibility: Both our research platform and our drafting assistant are free to use, no strings attached. We don’t do lock-ins and offer a generous free trial for anyone that wants to get a taste of the upgraded version. This removes all friction from trying the product out. We know once people use Iqidis the right way, they will not go back.

There’s a lot we’re working on, including real-time collaboration features, courses on how to use Legal AI, and more. Have something to say? Send me a message devansh@iqidis.ai. Working for social justice? Ask us about our pro-bono plans.

Making this Newsletter more Interactive.

It seems like a lot of you want more interactive stuff from the newsletter.

As I posted on the Group Chat, here are some of the things I'm playing with-

An interactive audio podcast link. You'll be able to listen to an audio podcast about the article breakdown. If you have any questions, you can pause the podcast and ask your question right there and it will answer for you. Will build this w/ LLMs + Voice AI.
Live streams. Thinking we'll do one of these every week. It'll be a bit more low-key, and we can come together to discuss AI related things, talk about papers etc.

The livestreams in particular seem to have drawn a lot of interest. Of these times, when seems like the best fit?

(anything else, message me. Provide times in EST please)

Streaming is not a space I know very well, so any inputs you can give me will be super helpful.

The Big LLM Architecture Comparison

Sebastian Raschka, PhD

needs no introduction. He’s an absolute legend in the AI space, and one of the best sources for technical AI Education. He dropped a comparison b/w LLMs and that triggered a question in my head.

Qwen3 works at 0.6B parameters. Kimi K2 works at 1T. That's a 1,666x difference. If intelligence needed a specific scale, this shouldn't be possible. But here we are.

What does that tell us about intelligence? If you ask me, this hints that intelligence isn’t related to scale or architecture. It’s related to pattern. Similar to fractals, and how they exist on different levels of zoom, but all follow the same structure. This wouldn’t be the first time fractals show up in NNs or encoding intelligence