Happy Monday! I spent this week keeping up with the latest AI / tech releases and updates so you don't have to. Here are the 4 things I found most interesting.
An unreleased Claude pushed a Riemann hypothesis bound from 41.6% to 67.2%, a number mathematicians had been inching up for decades.
On August 10, Anthropic published a result where a research version of Claude improved the lower bound for how many zeros of the Riemann zeta function satisfy the Riemann hypothesis, a problem that's been open since 1859. It took two runs: an earlier one that burned through around a thousand short-lived agents and got nowhere, then a second where Claude coordinated about 60 subagents over a day and a half, firing off 2,400 shell commands. Across both it used 31 million output tokens, and generated 650 ideas that didn't work before finding one that did. To be clear, this isn't a proof, it's a partial result, and it hasn't been peer reviewed yet.
My take: Doing all of this in two sessions sounds weird to me. It's impressive that they can kind of “brute-force” math problems now. Spawn however many subagents you can until something sticks.
Anthropic is sitting on an internal model that edges out its best released one, and it has no plans to ship it.
In its August 14 risk report, Anthropic disclosed an unreleased internal model it calls Model 2, describing it as more capable than Mythos 5 in some areas and less in others, overall slightly ahead. It says it has no plans to release the model externally and hasn't run its full suite of predeployment assessments, so it has lower confidence in what the thing can actually do. The same report raised Anthropic's risk rating for misalignment in high-stakes settings from "very low" to "low". They pinned that on general uncertainty after recent cybersecurity incident disclosures, not on anything the covered models did.
My take: Anthropic, the kings of marketing, back at it again. I respect the hustle though, and if they're not even confident about what this thing can do, no wonder they're sitting on it.
OpenAI previewed a mode that runs GPT-5.6 Sol at up to 750 tokens per second, roughly 14x faster than normal.
Announced August 13, Ultrafast mode is a new API tier that runs GPT-5.6 Sol on Cerebras wafer-scale chips instead of GPUs, keeping the model weights on-chip so there's a lot less waiting. OpenAI says it hits up to 750 output tokens per second, up to 14x its standard tier, which measures around 70 tokens per second at high effort. It's aimed at work where latency actually matters, like incident response, financial research, live customer support and commerce. It's in limited preview with a select group of customers, and there's no pricing or general release date yet.
My take: This is goated, but I'd never use GPT-5.6 Sol in my own projects. The pricing isn't that bad, it's just that if you're not doing anything revolutionary and are only building an AI app, these models are overkill. I hope this speed becomes the default though!
Google's Gemini 3.7 Flash is $0.75 per million input tokens until the end of the year, and it lands only three weeks after the last Flash.
Gemini 3.7 Flash dropped on August 13 across AI Studio, the Gemini API, Android Studio and Gemini Enterprise, aimed squarely at coding and agents. Against Gemini 3.6 Flash, which shipped on July 21, it jumped from 34.4% to 43.6% on FrontierCode 1.1 and from 49% to 65.3% on DeepSWE v1.1. It keeps the 1M token context window and lets you dial thinking between low, medium and high. Intro pricing is $0.75 / $3.75 per million tokens through December 31, then it doubles on January 1.
My take: Said it last time and it's still true, Flash models are the best value in AI right now. I'm fine with what I'm using but it's good to know the alternatives keep getting better. Love that Google is still improving the hell out of them.
On My Mind
Lately I've been consuming and reading a ton of “self-help” content. One of the things that has stuck with me is the idea of making the right choice. I graduated a month ago, and instead of going into a corporate job I had the privilege of pursuing content creation. If you look at it from the outside, the choice is incredibly clear. Imagine making 5x per month what you'd make going into a stable career. 1 great month of your content journey is equivalent to 5 months of working that job.
Yes, you have the opportunity to make much more and work for yourself. But that 5x isn't a salary, there's no floor under it. A bad month isn't just a bad month, it's a bad month with nothing promising you the next one. That's the part you don't see from the outside, and it's why you still feel the security under you slipping and still question if you made the right choice or not.
And the main thing I've learned is that there is no right choice. Life isn't a set path. You are responsible for making whatever choice you made THE right choice.
This is where understanding what you stand for, or what you desire, is important. I can't tell you if what I'm doing right now will still be working a year from now. That stability just isn't there, so I can't base my decision on it. What I do know is that I personally don't want to work under some big corporation. Startups are great and energetic, you usually own a part of the product and I am more than okay with that. It's always an open option 👀
But none of that gets decided today. What matters the most is your daily actions. The decision itself matters way less than what you do every single day, because those days are what define where you are in 3-6 months. So the questions I actually ask myself are simple. Am I improving the app today? Am I putting real work into my personal brand? Am I cutting down my bad habits? Those are the things that matter, and those actions are what make a decision a “correct” one.
So focus on your everyday actions, and I mean everyday. Look 3 months behind you. What have you been doing? Does it help you move towards your goals? If not, cut it or modify it until it does.
In life nothing is binary, there's no right or wrong path. The only thing I have is today and what I do with it. I'll find out in 3 months whether the small stuff actually added up, and honestly I'm fine either way.
More from me!
My last video: 3 hand-picked resources to learn the best language there is, Python! Watch here!
Tool of the week: The free voice tool I use every day to do all my work.
10x the context. Half the time.
Speak your prompts into ChatGPT or Claude and get detailed, paste-ready input that actually gives you useful output. Wispr Flow captures what you'd cut when typing. Free on Mac, Windows, and iPhone.
Have a great week, output more!
Volkan