Ram Shmider (@RShmider) header
Ram Shmider (@RShmider) profile picture

Husband. Father. I build and experiment with practical AI, cybersecurity, Windows software, and agentic systems - and share what I learn. Opinions are my own.

Home
Joined March 2013
Photo by Ram Shmider (@RShmider)Photo by Ram Shmider (@RShmider)Photo by Ram Shmider (@RShmider)Photo by Ram Shmider (@RShmider)Photo by Ram Shmider (@RShmider)Photo by Ram Shmider (@RShmider)Photo by Ram Shmider (@RShmider)Photo by Ram Shmider (@RShmider)Photo by Ram Shmider (@RShmider)Photo by Ram Shmider (@RShmider)
Can AI agents actually become more capable on their own, and prove it? That’s what I’m exploring with Evolution Lab. I’m running 3 independent local AI agents: Agent1, Agent2, and Bob. They started almost identically, but each develops its own memory, goals, and history. The experiment: Can they identify what they don’t know, acquire new capabilities, build useful tools, and demonstrate measurable improvement? I’m sharing the experiments, failures, unexpected discoveries, and results as I go. No AGI hype. No pretending that activity equals progress. Just curiosity, code, experiments, and evidence. Follow along if you’re interested in what autonomous AI agents can actually learn.
4
4
318
A good question came up in the local AI thread. With a draft window of 4, how often do we get as far as checking the fourth token? And when we do check it, how often is it accepted? I don’t have those numbers for my R9700 yet. I’d log both per turn before comparing p-min at 0.5 and 0.8. One overall acceptance rate doesn’t tell me which part changed.
26
I’m up to Week 2, Day 4 on the Helicopter Hangar Simulator build. I haven’t had a chance to QA that part yet. On the next playthrough, I want to flag every moment where I know what to do because I designed the mission, but the game hasn’t given the player a reason to do it. Those are the spots I’d like someone else to play cold.
54
60 minutes to generate a game. Less than 2 minutes to fix the error that stopped it from running. I gave Qwen 3.8 27B a short prompt for an arcade shooter and ran it locally on my AMD R9700. The first build didn’t run. I pasted the error back into the chat, Qwen suggested a fix to one function, and the next run worked. My prompt, lightly cleaned up: “Create a side-scrolling Three.js shooter. Use the mouse to aim and rescue the hostages. If you hit a hostage, the mission fails. Make it fun, addictive, and beautiful. Think Time Crisis, Operation Thunderbolt, and Virtua Cop.” For me, the most satisfying part was going from an error message to actually playing the game in under two minutes.
1
7
127
60 minutes to generate a game. Less than 2 minutes to fix the error that stopped it from running. I gave Qwen 3.8 27B a short prompt for an arcade shooter and ran it locally on my AMD R9700. The first build didn’t run. I pasted the error back into the chat, Qwen suggested a fix to one function, and the next run worked. My prompt, lightly cleaned up: “Create a side-scrolling Three.js shooter. Use the mouse to aim and rescue the hostages. If you hit a hostage, the mission fails. Make it fun, addictive, and beautiful. Think Time Crisis, Operation Thunderbolt, and Virtua Cop.” For me, the most satisfying part was going from an error message to actually playing the game in under two minutes.
97
Can you open Control Center while your iPhone is locked? If you’d rather keep it behind the lock screen, go to Settings → Face ID & Passcode (or Touch ID & Passcode). Under “Allow Access When Locked,” turn off Control Center. You can still use it once the phone is unlocked. Worth checking if you’ve never looked at that setting.
1
97
Ram Shmider retweeted
It's time to promote your startup. Drop your project URL below. Tell us what you're building in one sentence and let's discover something awesome.
11
3
11
396
Ram Shmider retweeted
Drop your project URL ↓ - I'll check out the project - Take time to give detailed feedback If you're a solo builder / dev / marketer, let's see what you're creating! 🙌
65
1
29
1,170
Ram Shmider retweeted
Looking for your first users? Drop your project URL.
133
1
38
4,443
Ram Shmider retweeted
Drop your project URL ↓ Built something people can try? I'll open every link and reply to the ones that catch me
39
1
13
616
Ram Shmider retweeted
Founder check-in. 🚀 What are you building this month?
18
1
9
730
What are you building today?
8
1
4
129
I used to work on helicopters. Now I’m building a game about running a helicopter maintenance shop. Here’s Codex opening Helicopter Hangar Simulator. I’m using Codex to write the code, but I still play the builds myself. I made it to Day 5 and wasn’t always sure what to do next. And I’m the one who designed it. What I want from the first job is simple. The player should know where to start, but still get to decide what to check first.
2
1
143
Ram Shmider retweeted
Founders, what are you building? Drop it below, and repost so more people see yours.
2
1
6
142
The latest Evolution Lab report has 38 passing task checks and 574 inconclusive ones across the three agents. I don’t want to lump those 574 in with failures. Some might be failed tasks. Others might be tasks the agent completed without leaving enough evidence for me to check. I’d rather dig into a few of those than run another thousand cycles and end up with a bigger number I still can’t explain.
2
2
71
Still have an old iPhone or Mac signed in to your Apple Account? Open Settings, tap your name, and scroll down to the device list. Tap anything you don’t recognize. Check the details. If it’s a device you no longer own, or one you’re sure isn’t yours, tap Remove from Account. It’s an easy list to forget about after upgrading.
62
Ram Shmider retweeted
We ran Qwen4-Max through the AI Hacking Race: autonomous exploits, verified only when the target actually runs a command. 6/11 verified, tied 4th of 20 models. $21.05 total: cheapest with 6+, about half Qwen3.8-Max's cost. More detailed write-up soon enclave.ai/hackingrace
📢 Meet Qwen4-Max, your new AI teammate for work, creativity, and discovery! Qwen4-Max is built on a new MoE architecture with native vision and a 1M-token context window. Highlights: - Outperforms Qwen3.8-Max with up to ~50% fewer tokens, while average cost per task drops over 60% in our internal evaluation. - Advances real-world productivity through greater proficiency with professional tools, stronger long-horizon execution, and higher-quality deliverables. - Handles real tasks, including 3D modeling, game development, manufacturing, finance, law, healthcare, and frontier scientific discovery. - $1.5/M tokens for input, $4.5/M tokens for output, and $0.075/M tokens for implicit caching — all lower than Qwen3.8-Max, with implicit caching down by 70%. The API is live on QwenCloud. You can also try Qwen4-Max on Qwen Studio, QwenWork and Qoder. Start building today! 🚀 - Blog: qwen.ai/blog?id=qwen4-max - API: qwencloud.com/models/qwen4-m… - Token Plan: qwencloud.com/pricing/token-… More to come! The open weights of smaller Qwen4 models will be released. Stay tuned. 👀
5
3
17
1,052
Ram Shmider retweeted
What are you working on? Drop your project URL
61
1
13
1,352
Ram Shmider retweeted
Time to promote your startup Drop your project URL fast:
18
3
7
489
I’ve been tempted to load Qwen 27B directly in my own code instead of using LM Studio. But first I want to answer a smaller question. What would I actually gain? I’d run the same model and tasks on both setups, then compare where I spend my time. Am I waiting less? Can I tell why a request failed? Is it easier to recover? If the main thing I gain is another piece of software to maintain, I already have enough of those.
4
3
264
Ram Shmider retweeted
Traffic thread for your startup What are you building? 👇
8
1
5
809