Mateusz Mirkowski

My website redesign made with Gemini. Kinda cool I think. 😇 Not finished yet. #ai #gemini

@Logos_devs @0xSero True.

If you're an AI power user, get prepared. As @0xSero says, the math for subs doesn't add up. $200/mo can't cover the 16+ GPUs needed for power users. So what happens next? → Prices go up → Limits get tighter → Your data becomes the product But there’s another path emerging. Open-source models are coming out with incredible efficiency savings. Local inference is improving. Home hardware is becoming viable. The frontier is shifting from access → ownership. Start building your sovereignty now.

Don't learn to code. Coding is obsolete skill. Learn Software Engineering and AI.

Play games with AI? Why not. ;)

@Pawzgm Great job! Maybe I will try it on 4090.

I refused to accept 35 t/s on Qwen3.5 27B. Made a custom quant. Got 45. Same eval scores as the unquantized model. If you want to try it yourself: huggingface.co/Pawellll/Qwen3… Would love to hear your numbers, especially on smaller GPUs.

Playing for the first time. 🥰

@pdrmnvd Stop reading most parts. Verify features by tests, AI code review, use cases, documentations. Read only critical parts.

my ability to produce code is exceeding my ability to read and understand it and i have no idea what to do about it.

@bridgebench This is why I am building my claw. :)

@visualdevguy ha, queue it up before bed and check the results in the morning. not the worst strategy

GLM 5.1 is the slowest frontier model we've ever benchmarked on BridgeBench. 44.3 tokens per second. Half the speed of GPT 5.4. Nearly 6x slower than Grok 4.20. Z.ai traded all of their speed for intelligence. The coding benchmarks improved. The throughput collapsed. In 2026, agentic coding is about parallelism. You're running 5, 10, 15 agents at once. A model this slow bottlenecks every workflow it touches. Intelligence without speed is a luxury most vibe coders can't afford. bridgebench.ai

@ivanfioravanti On weekend I will test it on my 4090. I wonder what difference will be compare to mac 64gb.

Don't tell anyone 👀, but Unsloth Qwen3.5-27B-UD-Q4_K_XL on Nvidia 3090 TI is a real beast. M5 Max is amazing and fast with MoE models, but with dense model there is no match. With MLX performance are much better, but 3090 TI wins anyway. 🤷🏻‍♂️ My Hermes Agent loves this model!

@bridgemindai Yup, and I fell like limits are lower than after gpt 5.4 release.

@visualdevguy that's impressive for $20. Plus plan is solid value

Been running Codex with GPT 5.4 High all morning. 87% of my 5 hour limit still left. On Claude Code I'd be at 0% by now. OpenAI gives you limits you can actually work with. I've been shipping all day without a single rate limit or 529 error. This is what $200/month gets you on Codex.

@Presidentlin Nice. Off-peak time is my morning, working time. <3

Model is up on OpenCode btw. I think this is going to be my lineup Opus: GLM 5.1 Sonnet: GLM 4.7 Haiku: GLM 4.7-FlashX The GLM 5 series Consumes quota at a rate of 3× during peak hours and 2× during off-peak hours. They have a special "As a limited-time benefit, GLM-5.1 and GLM-5-Turbo will count as 1× during off-peak hours until the end of April. Peak hours are from 14:00 to 18:00 (UTC+8) daily." So for me I shoud not be using them for most of the morning after 8 am. Which is fine, my order will be: Google OAI Zai OAI is mostly for big-brain tasks or attention to detail. Google with 3.1 Flash Lite and 3.1 Flash are for spamming. Zai will be for when Google is in cool down. This is me being spoiled, really, my time is best spent marketing my stuff, then churning code.

@scaling01 GPT 5.4 is not smarter than Opus 4.6. It's just more reliable.

I've actually resubscribed to ChatGPT Plus because I'm currently building a lot, and a $200 subscription is too much I rather get Claude Pro and ChatGPT Plus for $40 what I've noticed: - GPT-5.4 is smarter than Opus 4.6 - Opus and Sonnet make such silly mistakes sometimes and I don't feel like they are actually thinking about what they are doing. They still feel like they are exploitation maxxing

@hxiao Very good insight. Thanks.

Rising interest in Qwen3.5-27B dense over 35B-A3B. If you're running on a budget GPU like L4 24GB, you need to know that 35B-A3B is still 7x faster decode & 4x longer context at identical quality: 256K vs 71K in the same 24GB vram, making 35B-A3B the better choice for long context tasks like synthetic data gen.

@_devJNS 2025 so much cleaner and better!

lmfaoo 😭

@design_nocodeio I heard many times about Railway, why it's popular?

Working on a project for myself this week. Stretching FlutterFlow & Claude code to help with my backend. All of this is live data -- coming from API's a custom backend built entirely with Claude and running on Railway.

@diegocabezas01 Not for people, only for elites.

Anthropic Claude Mythos = Capybara

@LottoLabs Outstanding. I have 4090 and m1 max 64gb. I plan to buy Mac studio 128gb. I think I should buy it ASAP because hardware prices gonna spike soon.

@visualdevguy Yeah this is just the last week of usage lol

Qwen 27b on the 3090 saving me a bag. This is cost savings for 7 days of usage, w/ Hermes agent. Assuming 80% cache hit (unlikely) and no cache timeout. This is conservative. 27b is between sonnet and 5.4 mini This is just my tokens in/out w/ api costs, assuming no rate limits. Obviously cheaper w/ coding plans $200/m but would be hitting limits likely.

That's why local llms are winning in 2026. If you don't have mac mini/studio, buy it now, because hardware gonna spike soon.