codvyn (@codvynn) header
codvyn (@codvynn) profile picture

260k on IG | Tech & AI Helping Founders Automate everything.

Joined January 2026
Photo by codvyn (@codvynn)Photo by codvyn (@codvynn)Photo by codvyn (@codvynn)
Daily AI updates | Day 42 MAI Code 2: Microsoft's new coder claims 98% of GPT-6.1 Sol's score at 40% of the price, 96% of Opus 5.5 at 5%. Optimizely's CDN served malware to 178 orgs on Nov 16, Zoom and GitLab included. The payload came through customers' own project JS, and kept Optimizely's certificate. Windows 11 logs which coding agents you run (codex, claude, cursor) and sends it up unless you turn diagnostics off.
1
54
Carnegie Mellon is publishing 8 of its CS courses on YouTube, free. 11-768 AI Agents, 15-442 Machine Learning Systems, 17-645 ML in Production, 10-605 ML with Large Datasets. Faculty teaching their own classes, 5 more in January. cs.cmu.edu/learn/
1
3
112
What makes you trust an AI-written PR? Mine has to survive me explaining it out loud.
4
6
167
64 A characters and one equals sign corrupted the heap in OpenJDK's Base64 decoder. CVE-2026-114938, CVSS 9.2. Oracle patched Nov 17. Can't update yet? -XX:-UseBASE64Intrinsics
2
72
What's the one thing you wish your editor could do that nobody has built yet?
1
4
119
An open-source agent went from first scan to full compromise in under three hours, per the FBI. The same tool hit 2,250 ConnectWise customers and 68,000 people in South Korea.
1
62
Day 10/30 Looking for Builders, Founders and People making cool stuff! Drop your SaaS, AI tools, Side projects, anything :) Let's #connect 🤝 #buildinpublic
63
33
986
What would you build if deploys took 30 seconds instead of 10 minutes?
1
4
112
Daily AI updates | Day 41 Microsoft Research and UC Berkeley open-sourced TokenEconomizer, a plugin that learns from past sessions in one repo. On 277 SWE-Bench tasks it cut cost 18.4-30.9%, cheaper in all 30 repo and model pairs, success up in four of five. The waste is boring: 23 tool descriptions sent per turn, 16 never called, and the 17% of tasks that failed ate 32% of the spend. Windows 11 matches codex.exe, claude.exe, cursor.exe, copilot.exe and five more inside MissionControlAggregator.dll, feature CLICodingAgentUsage. With optional diagnostics on, that usage data goes to Microsoft. Optimizely's Project JavaScript was hijacked on 16 Nov and served from its own CDN to 178 organisations, Zoom and GitLab among them. 169 snippets were infected at 20:56 UTC. One at 21:52.
2
84
How many lines of AI-written code did you actually read line by line this week?
4
4
108
Mapbox charges $5 per 1,000 map loads once you pass the free tier. A single 139GB file in your own storage bucket covers the whole world, served over HTTP range requests, at roughly $0.05 per 1,000 loads. No tile server. No per-load API key. docs.protomaps.com/
1
1
78
What's your rule for when to stop planning and start building? Mine is when the plan gets longer than the code.
2
4
82
Name one thing that got slower the moment you started using AI for it.
1
39
Four open-source tools for the boring half of shipping AI features: 1. promptfoo: test prompts and agents, compare models 2. litellm: one OpenAI-shaped call for 100+ models, with cost tracking 3. langfuse: a trace per request, see where the money went 4. tiktoken: count tokens before you send, not after the bill None of these make the demo better. All of them keep it from getting expensive.
6
6
174
Your agent breaks prod. Do you revert it, or let it fix its own mistake?
1
36
The code nobody wants to touch is usually the code that pays the bills. Everyone knows where it is. Everyone routes around it. The scary part is that the reasons left with the person who wrote it. Who is the one person your stack still depends on?
1
44
What's still running in production that you'd rather not touch? You know the one. It works, and nobody remembers why.
38
Day 9/30 Looking for Builders, Founders and People making cool stuff! Drop your SaaS, AI tools, Side projects, anything :) Let's #connect 🤝 #buildinpublic
49
1
21
777
What's the tell that a benchmark table is cherry-picked? Mine: the missing baseline. The model it should beat isn't in the chart.
2
3
90