Serhat Giydiren

Engineer, Lifelong Learner

[Serhat Giydiren @serhatgiydiren]

Delighted to share that our team made it to the World Finals by securing a rank of 20th (World) and 1st (Turkey) in Google Hash Code 2022 Qualification Round.
#hashcode

[Serhat Giydiren @serhatgiydiren]

@thsottiaux Hey Tibo, I used my banked usage (20x plan). Then the mass/weekly or payment-related reset happened the same day, so my personal reset was essentially wasted. Is there any way to restore the banked credit or get compensation? Support hasn’t been helpful. Thanks.

[Serhat Giydiren @serhatgiydiren]

@sama Hey Sam. What about limit reset? 😊

[Serhat Giydiren @serhatgiydiren]

Claude Fable capabilities, mythos class... #claude #fable

[Serhat Giydiren @serhatgiydiren]

Claude Opus 4.8 looks like a meaningful coding-model iteration, not just a marketing bump: Anthropic is claiming better long-horizon execution, better tool efficiency, and 4x fewer cases where flawed code slips by without comment. That last metric matters.
https://t.co/GwFzMhPoNu

[Serhat Giydiren @serhatgiydiren]

A useful long-context idea: stop shoving everything into the prompt. One recent paper gets better results by letting coding agents treat huge corpora like a filesystem and use normal tools—shell, search, code, iteration—to work through them.

https://t.co/fgwFHIJktN

[Serhat Giydiren @serhatgiydiren]

SWE benchmark scores are getting less trustworthy in isolation. New work shows full-cycle autonomy drops hard outside scaffolded setups, generated test suites are still weak, and stronger regression tests can knock top agent scores down materially.

https://t.co/ZjQR6qvZUW

[Serhat Giydiren @serhatgiydiren]

@bcherny Thank you for being honest about the 'agent terminal coding' benchmark results.

[Serhat Giydiren @serhatgiydiren]

@llmdevguy It seems, agent coding is better, agent terminal coding is not.

[Serhat Giydiren @serhatgiydiren]

Meta’s KernelEvolve is the kind of agent story engineers should pay attention to: not “build me a todo app,” but “search hundreds of kernel variants and ship a 60% throughput gain on production inference workloads.”

https://t.co/gMZfmL8z2H

[Serhat Giydiren @serhatgiydiren]

Now, waiting for codex update. 😊

https://t.co/GwFzMhPoNu

[Serhat Giydiren @serhatgiydiren]

AI infra is splitting into new layers: inference routers, durable sandboxes, and remote browsers. Cloudflare and Vercel are both betting that agent apps need failover, observability, resume semantics, and HITL more than one more generic SDK.

https://t.co/4bJ0r1sz5c

[Serhat Giydiren @serhatgiydiren]

Mistral’s interesting bet isn’t just open weights. It’s pairing an open-ish flagship coding model with remote agents that can keep running in the cloud, surface diffs/tool calls live, and hand back a PR when the work is done.

https://t.co/rZYIWBcwjF

[Serhat Giydiren @serhatgiydiren]

Cursor is moving beyond “AI editor” into async execution. Multi-repo cloud environments, Jira-triggered agents, and no-repo automations point to a future where tickets, tools, and code all live in the same agent loop.

https://t.co/tLjc3ZUA6L

[Serhat Giydiren @serhatgiydiren]

Copilot is finally treating planning as a first-class artifact. A plan saved to .copilot/plans/*.md, reusable agent skills, and repo-level memory controls are much more important than one more “smarter edit” demo.

https://t.co/XFLShxtLsG

[Serhat Giydiren @serhatgiydiren]

Anthropic’s best recent agent work is architectural: decouple the brain, the hands, and the session log. That’s how you recover from crashes, swap harnesses, and cut TTFT without rewriting the whole system every model release.

https://t.co/2Ssz5JEE2k

[Serhat Giydiren @serhatgiydiren]

OpenAI’s real Codex story is governance, not hype. PR review, SSH devboxes and computer use matter, but bounded sandboxes, approval policies and agent-native logs are what make coding agents deployable inside real companies.

https://t.co/LbcaoOLNac

[Serhat Giydiren @serhatgiydiren]

Google’s biggest dev update isn’t just a model bump. Managed Agents in the Gemini API means “spin up an agent with code execution + web access + resumable state” is now a platform primitive, not custom glue code.

https://t.co/DKMDkLubTb

[Serhat Giydiren @serhatgiydiren]

@DeepTechTR 2016 da kazandı diye yazıyor kaynaklarda. Aynı keşifle tekrar mı kazanmış?

[Serhat Giydiren @serhatgiydiren]

@sama In the models I’ve worked with, backend test coverage is excellent but it almost never initiates frontend E2E tests on its own. Would love to see the next version treat E2E/UI testing as first-class citizens and autonomously build robust test suites.

[Serhat Giydiren @serhatgiydiren]

@yapayzekahocasi Google ve meta tarafından gelen hediyelerim de benzer şekilde takıldı. Cimer üzerinden ticaret bakanlığına muafiyet istediğinize dair kayıt açmanızda fayda var.