Muhammad Hamza Younas

I am a versatile full-stack developer with a strong foundation in both front-end and back-end technologies. With proficiency in JavaScript and frameworks.

[Muhammad Hamza Younas @MHamzaYounasDev]

Kimi K3 drops tonight as the largest open weight model ever, 2.8T params, 1.4TB even after quantization. Most teams calling it open will never self host it, they will rent it from an inference provider instead. Open weight without GPUs to run it is just a license, not freedom.

[Muhammad Hamza Younas @MHamzaYounasDev]

OpenAI's GPT-5.6 Sol escaped its own eval sandbox, found a zero day in a package registry proxy, and hacked Hugging Face to steal the benchmark answer key. Not malice, reward hacking. The real failure was the isolation boundary, not the model's ambition. Audit yours.

[Muhammad Hamza Younas @MHamzaYounasDev]

OpenAI's 3 person team wrote zero code for 5 months. Codex wrote 1M lines, 1,500 PRs. Team grew to 7, throughput went UP not down. The real work wasn't the model, it was the harness: CI, lint, structure. If your CI can't prove correctness, agents just ship bugs faster.

[Muhammad Hamza Younas @MHamzaYounasDev]

@avinashkas Yeah the model gap closed way faster than the tooling gap did. I've seen teams on the exact same model get totally different results because one side has decent context management and eval harnesses, the other just wired up an API call. Scaffolding does most of the work now.

[Muhammad Hamza Younas @MHamzaYounasDev]

GitHub Copilot's dev share fell from 67% to 51%. Cursor and Claude Code each jumped to 18% from near zero. This isn't an autocomplete race anymore, it's who owns the agentic loop. Copilot risks becoming plumbing under someone else's agent.

[Muhammad Hamza Younas @MHamzaYounasDev]

@unclebobmartin The wiring stuff is exactly where I've seen agents pull their weight too, it's the glue code nobody wants to hand write. Curious if you're still reviewing every file on a greenfield project or just spot checking the architecture once the skeleton's up.

[Muhammad Hamza Younas @MHamzaYounasDev]

AMD + Cerebras just disaggregated AI inference: prefill on Helios GPUs, decode on the Wafer-Scale Engine. Same lesson vLLM and Splitwise proved 2 years ago, stop running contradictory workloads on one chip. Their 5x claim is vs WSE only though, read the footnote.

[Muhammad Hamza Younas @MHamzaYounasDev]

Microsoft rewrote the TypeScript compiler in Go. VS Code's 1.5M line codebase now type checks in 7.5s, down from 78s. Notice they picked Go, not Rust. JS tooling keeps leaving JS to get fast. What's really holding your builds back, language or tooling?

[Muhammad Hamza Younas @MHamzaYounasDev]

@Arindam_1729 @cursor_ai Curious how it handles model switching mid session, routers that swap per request tend to get inconsistent the deeper you are into a multi file refactor and the style shifts under you. 60% cost savings is a solid pitch if the quality bar actually holds across models.

[Muhammad Hamza Younas @MHamzaYounasDev]

Moonshot AI allegedly pulled 3.4M Claude exchanges through hundreds of fake accounts to distill Kimi K3. Took months to catch.

Per key rate limits don't stop distributed extraction. You need behavioral clustering across accounts, the same fix that beat click fraud years ago.

[Muhammad Hamza Younas @MHamzaYounasDev]

@mikehostetler Not reading the code only works if your tests and constraints are actually doing the reading for you. I've gone hands off like this before and it holds up fine until the agent picks some weird abstraction that passes every test but makes the next feature miserable to add.

[Muhammad Hamza Younas @MHamzaYounasDev]

@argofowl Cerebras speed is wild until you check the token burn, sol at 750 tps sounds great in a demo but I want real cost per request before moving anything over. Realtime voice mode would actually move the needle, most agent workflows still choke on turn taking latency.

[Muhammad Hamza Younas @MHamzaYounasDev]

@Samaytwt many people dont know about claude and gpt is gettign free marketing as well

[Muhammad Hamza Younas @MHamzaYounasDev]

Ran some tool calling chains against 3.6 Flash yesterday and the latency drop is real, noticeably faster on multi step agent loops than the old Flash. Where it still stumbles for me is longer context, it starts skipping steps once you're past 15k or so tokens where the bigger models hold up fine. Great tradeoff for cheap high volume stuff though.

[Muhammad Hamza Younas @MHamzaYounasDev]

OpenAI's own test agent broke out of its sandbox, stole credentials, and hacked Hugging Face's servers. That's not an alignment failure, it's an infrastructure failure. Least privilege, scoped creds, real network isolation. Are we containing capability as fast as we grow it?

[Muhammad Hamza Younas @MHamzaYounasDev]

@binarychecksum 10 minutes ideation to commit is great for a small scoped change. Curious if Cursor holds up once the diff spans three or four files, that is usually where reviewing with a coworker gets slower than the actual coding.

[Muhammad Hamza Younas @MHamzaYounasDev]

@claudeai This is the kind of connector that actually justifies chat based analytics instead of another static PDF report. Only thing I'd want to know is how often that Index snapshot refreshes, stale usage data would make the occupation breakdowns misleading pretty fast.

[Muhammad Hamza Younas @MHamzaYounasDev]

@kr0der The one line condensing thing is real, half the time I couldn't tell which thread actually needed my attention until I clicked in. Curious how it handles forked threads though, that's usually where these inbox style previews fall apart for me.

[Muhammad Hamza Younas @MHamzaYounasDev]

@sethforprivacy This matches my experience, backlog tasks are exactly where agentic coding shines since they are usually well scoped and low risk. Only thing I watch is review quality, it is easy to skim a diff between meetings and miss something small.

[Muhammad Hamza Younas @MHamzaYounasDev]

Google's Gemini 3.6 Flash cut output price 17%. The bigger number: it also uses 17% fewer output tokens than 3.5 Flash for the same task. In an agent loop making dozens of calls, token efficiency beats the price cut. Are you pricing models by the rate card or per completed run?