Elijah Popowitz

We achieved SOTA on Terminal Bench 2.1 using a DeepSeek v4.1 Flash model. 92.2% at $0.08 cost per task. Previously Opus 5.5 was the leader at 87.6% at $0.50 cost per task (vals.ai/benchmarks/terminal-…). Now the obvious question is how did we do it? How can you possibly improve a Flash model to beat frontier models like Opus, Fable, and Astra? The truth is, we over-obsess about models when they are only half of the equation. The other half is the code that wraps around the model. In fact all of the most revolutionary products (Instinct, Muse, Claude Code, OpenClaw, etc.) are all just harnesses optimizations. Rather than try to build a new model or fine-tune, we built a harness that would make: Flash Models —> Frontier Results Frontier Models —> Beyond Frontier Results Did I mention it’s also ~6x faster than Claude Code? And we’re now live. Go Check it out! Trymidas.io

Thank you everyone for the support! Since this post reached 50k views, next 50 people whose first deposit is >$10 get 50% bonus credits (Up to $50 bonus credits) Sign up with this link: trymidas.io/r/LAUNCH50 or In Dashboard, go to Billing Code: LAUNCH50 $10 deposit = $15 in credits ... $100 deposit = $150 in credits

We also achieved SOTA on @CommandCodeAI's Frontier Token Efficiency eval. This benchmark measures read and shell tool calling token-efficiency and we crushed. Read more: trymidas.io/benchmarks/token… (Yes we will run TB4 in the next couple weeks) Since this post went viral, next 50 people whose first deposit is >$10 get 50% bonus credits up to $50. $10 deposit = $15 in credits ... $100 deposit = $150 in credits In Dashboard, go to Billing Code: LAUNCH50

My perspective on groupthink and consensus. View article

The Knowledge Fallacy The knowledge fallacy is an anecdote about questioning what you hear and being more curious. This essay is not politically driven but rather a broader theme that I’ve witnessed play out over the past few years. As humanity’s access to information has grown, we have become innately less curious. This sounds counterintuitive, but it’s real and not necessarily for the reasons you might think. Attention: We have a limited amount of attention, and in the digital age we are overstimulated by a constant influx of information. Whether it be war, famine, tragedy, what our political leaders recently said, or the most recent AI scare, we are unable to give the appropriate amount of attention to any topic because there is always something new to distract us. Our brains are not designed to be this stimulated. Before the internet, our perception of the world was merely what we experienced in our lives (basically what happened in front of us and whatever was on satellite TV). In the digital age, however, we are always just one notification or one swipe away from a new message, oftentimes someone trying to persuade us or sell us something. The Result: 1. We only read headlines and take them at face value, which leads to clickbait and the spreading of misinformation. 2. Every company and creator is fighting for the same finite attention, benefiting from short-form content that has rotted our attention spans. 3. We seek stimulation and amazement like the prestige at a magic show. Everything has to be bold or outlandish to deserve our attention. 4. The short-term beneficiaries of this phenomenon are what pop culture now calls “larpers.” They attract attention by showing some outlandish demonstration but if you were to look deeper, there is little to no substance behind their claim. Everything is about misdirection and fooling the consumer. This is only accentuated by the emergence of AI, which has decreased the amount of effort that it takes to produce content or a product, allowing the release of endless “slop” into the world. Consumers recognize this and resort to “trusted sources” or “experts,” but this is where it gets particularly dangerous. The Danger: Industry leaders and public figures who are deemed trusted sources have an inherent responsibility to not mislead the public. The issue with that is people with large enough audiences have enough influence to sway public opinion and will oftentimes do so to perpetuate their own agendas. Their messages create a level of groupthink that hasn’t been present since the days where information was restricted. Groupthink is possibly the most dangerous outcome of all. People will adopt opinions from experts that are not their own and parrot them to others to demonstrate their own intelligence. The worst part is they don’t even know they’re doing it. We live in a world that rewards perception or appearance above all else and that’s why people will make claims or have opinions about subjects they know nothing about. This goes for politics, the perception of a startup, an industry, a person, or whatever else. My Own Experience with Groupthink: I am a startup founder, building in a space that I’ve been told a million times is too crowded, that the market saturated, and that my competitors have too much talent density that I can’t possibly win (whatever the fuck that means). The truth is that most of the people telling me these things have never used a product in my space. They haven’t done real research, they don’t understand the intricacies of what these products do, and they’ve fallen for groupthink because they read a headline. Maybe that headline was a revenue number or a big financing round, and that signalmade them so confident that the market has already been won, that they didn’t do their own research and come up with their own conclusion. We were able to close a 7-figure pilot in our first week. My ambitions aside, this is a dangerous reality we’re seeing a lot in Silicon Valley for founders and investors alike. An industry leader will make claims such as: Software is dead. Companies will build their own SaaS. Coding is solved. A vocal minority will adopt their opinions and these ideas will spread. I use these examples because all three were claimed and disproven in the past year. But this goes back to my initial point about clickbait and why we shouldn’t blindly listen to leaders, because they aren’t always right. What Should You Take Away From This? Continue to respect the opinions of others who are knowledgeable but do not treat their words as gospel. Make sure that you do your own research and come to your own conclusions before blindly adopting the opinions of people who are trying to persuade you. Take back your sense of individuality and try to discern what is real and what is noise. See how I just convinced you? My point exactly.

The craziest part about the 76ers is that they have 5-6 guys who can average 25+. The only thing stopping them from dropping 150 points a game is that there’s only one ball

Who wants free AI credits?

I’m looking for beta testers

At what stage do startups move from subscription to usage based for Codex/Claude?

I built my own coding agent... And then I cancelled my Claude & Codex subscriptions

How long do you think it will take the general public to adopt AI the same way tech/Silicon Valley have?

In light of the “Pace the Frontier” movement, I think this is a particularly good time to write my opinion about the “AI bubble” and how I believe many people misunderstand it. Context This past weekend, Dario Amodei (founder/CEO of Anthropic) called for regulation of AI and a more calculated/safe approach to the development of future models as they race to achieve super-intelligence, more commonly referred to as AGI (where models become smarter than people). In the past few quarters, model development has undergone a series of breakthroughs that have caused a distinct ideological change from chasing benchmarks towards creating a super-intelligent artificial entity. Theoretically, they’re building towards a model that, with enough resources (compute), would have the ability to solve humanity’s most pressing issues: abundant energy, space travel, cures to complex diseases, etc. What I find particularly fascinating about this phenomenon is how it alters the AI landscape and the (largely uninformed) public’s perspective on this emerging technology. AI Safety I plan to write another essay just on this subject but for the sake of this read, I’ll keep this somewhat brief. Without sounding too AGI-pilled, AI has the potential to become one of the most pivotal creations in human history: democratizing access to information, curing terminal diseases, and allowing for a wave of entrepreneurship because it permits a single person to accomplish what used to take hundreds of people to do. But like any great technology, used incorrectly, it can cause significant damage (no different than the internet, social media, or nuclear energy). Do I think AI is going to kill us by 2030? No. But that said, I think it’s important that we start having these conversations now and we consider what the future might look like with these advanced models. In particular, we need to ensure that this technology benefits humanity as a whole rather than just satisfying the monetary benefits of the few. Now to the fun stuff! Is this Really a Bubble? Many see the high valuations from the frontier labs that are creating relatively minimal economic impact (so far) and automatically assume that we’re in a bubble. I once thought the same. I think it comes from the misconception that AI = Automation or AI = Magic (many people’s first response to using ChatGPT). The reality is that this divergence between technology and ROI comes from a discrepancy between adoption (what people use AI for) and what the labs are building. As model capabilities improve, the vast majority of consumers continue to use these models the same way they used models from last year: advanced Google searches, brainstorming tools, cooking recipes, etc. Therefore, we can draw many similarities between the “AI bubble” and the “Dot Com bubble.” The issue isn’t the underlying technology itself but a mismatch of that technology in relation to its supposed ICP (ideal customer profile). I believe the way to view this mismatch is simply a difference in motives. Frontier labs are training with the goal of creating more intelligent models, whereas the vast majority of businesses and people aren’t looking for intelligence; they are looking for effective task completion. So when they pay for increased intelligence and receive marginally beneficial results, they’re disappointed and claim the technology has plateaued. Conversely, there is a small subset of people who are receiving tremendous value from these new Fable-level frontier models, specifically experts working at the top of their fields. This is an important distinction. AI has historically been a floor raiser, allowing the average person to execute at an intermediate level in a domain they have less experience in. Experts, by that same logic, didn’t benefit because AI outputs were lower quality than what an expert could produce on their own. With Fable-level models, they can operate as ceiling raisers when used by someone with strong domain expertise. This doesn’t mean that everyone can now perform at this expert level (AI continues to behave as a floor raiser for the average person), but rather that experts are the primary beneficiaries of this increased intelligence. In sum, my argument isn’t that the frontier labs will crash and burn, but rather that AI models will be sorted into at least two different buckets. 1. Super-intelligent models that, with brute force, can solve problems that humans can’t solve alone (Frontier labs). 2. Highly capable AI models that exist to automate repetitive tasks for the majority of consumers and businesses (similar to TypeSafe’s philosophy). Capability vs Intelligence While these terms seem synonymous, they are in fact fundamentally different. Take Albert Einstein for example. He is a genius. Ask Einstein to operate an iPhone and he would be less capable than an average 12-year-old. It’s not an intelligence issue; Einstein simply hasn’t used an iPhone before. Now give Einstein 6 months of training and he would perform similarly to the 12-year-old. This means intelligence can translate to capability, but if you needed tech support, you still wouldn’t call for Einstein. That’s the problem with the commercialization of frontier models today. Most consumers and businesses don’t need that much horsepower to complete daily tasks; they just need those tasks done correctly. So Where Does This Leave Us? Superior intelligence is something that we can strive for, but it’s unlikely that these advanced models will be used to complete the majority of agentic task work. TypeSafe has the right idea, but their current model is too static. In my opinion, the future of mass-distributed AI will be self-learning and customized on a per-organization basis, specifically optimizing for task efficiency. (Note: Companies will not be able to do this alone, just as they didn’t successfully build their own SaaS). Therefore, the commercially successful labs of the future will release base models into organizations that learn from their surroundings and interactions, form neural pathways (adjusting weights), and grow into the efficient workhorses that companies are looking for, no different from how a new employee would. This is the model we’re building at Midas.