Kıvanç Yüksel

One-person team behind SmileToAI. I ship features, read robotics textbooks, and program my own hypertrophy blocks. Warsaw.

[Kıvanç Yüksel @thejoyrises]

https://t.co/CuBjGETzcj

[Kıvanç Yüksel @thejoyrises]

pgvector applies your WHERE after the HNSW scan. Per its README, a filter matching 10% of rows returns ~4 results at the default ef_search of 40.

I checked our notebook search: one HNSW index over every team's chunks, team and notebook in the WHERE, iterative scan never enabled.

[Kıvanç Yüksel @thejoyrises]

@DvirDu Thanks, we're sorted. It was our own change, we'd lowered NODE_CREATION_BUFFER below what the graph was built with. Only thing worth an issue: it failed silently. A startup warning for that would have caught it.

[Kıvanç Yüksel @thejoyrises]

Our graph database went two days without writing a snapshot, and Redis said the last save was ok the whole time.

We had dropped FalkorDB NODE_CREATION_BUFFER from 16384 to 1024. Every background save failed, 866,881 of them.

Alert on rdb_last_save_time age, not the status.

[Kıvanç Yüksel @thejoyrises]

@loudchirper You drew the line at residency, per-task cost caps, outbound limits and multi-model audits. Nothing about a cheaper tier implies those, so downmarket closes the window for consumer wrappers, not the plumbing. Unless the bet is staff bringing it in on the cheap tier.

[Kıvanç Yüksel @thejoyrises]

@ThorCommerce Fair. A fresh rdb_last_save_time only proves the write finished, not that the file loads. A throwaway restore plus a check on a recent node is going on the list.

[Kıvanç Yüksel @thejoyrises]

https://t.co/r6PVHUPtl6

[Kıvanç Yüksel @thejoyrises]

Gemini 4 Argon's output limit went from 64K to 1M tokens. At Google's post-intro price ($20 per 1M output), one call that uses all of it is $20.

One of our pipelines sends no output cap on purpose and takes the model's max. That was fine at 64K. I wouldn't leave it unset at 1M.

[Kıvanç Yüksel @thejoyrises]

Same prompt, 4 models #1: all four spelled my 3-line sign right on the first try. Two also painted signs I never asked for.

Nano Banana 2 added a farm banner and two produce boards, GPT Image 2 a LOCAL HONEY sign. Grok Imagine and FLUX.2 Pro didn't. One run each.

[Kıvanç Yüksel @thejoyrises]

https://t.co/lj3LpiD0sd

[Kıvanç Yüksel @thejoyrises]

A red scarf and a green scarf can turn into the exact same grey pixel: RGB(196,52,52) and RGB(30,140,30) both convert to 95.

So a colorizer picks a color that fits. It can't get the real one back. In our colorize guide I wrote "plausible, not documented" instead of "accurate".

[Kıvanç Yüksel @thejoyrises]

https://t.co/sPOJWyNnI4

[Kıvanç Yüksel @thejoyrises]

Opus 5.5 is 20% cheaper per token than Opus 5, $4/$20 per million in/out vs $5/$25.

The part I'd read twice: thinking can't be turned off. thinking: disabled is a 400, and thinking tokens bill as output. A job that ran with thinking off can end up costing more per request.

[Kıvanç Yüksel @thejoyrises]

@oldstackjournal The loud breakages are the fun ones, you get your Saturday and your three lessons. What got me was a quiet one: my self-hosted scheduler failed a post and left it out of its own API listing, so my monitoring said all clear.

[Kıvanç Yüksel @thejoyrises]

@nutlope @ollama What I'd want in the benchmarks is calibration against Jev. My cutoffs were tuned on Jev's scores, so swapping in a local model only works if 0.5 still means roughly the same thing on Tev1. Accuracy alone won't tell you that.

[Kıvanç Yüksel @thejoyrises]

An image model drew garbled bullet text into an infographic, the exact thing my prompt said not to draw. Re-ran that prompt 3 times, all clean. Rewrote it, 4 more renders, also clean.

I'm not calling that fixed. A failure that rare barely shows up in a handful of runs.

[Kıvanç Yüksel @thejoyrises]

@JayaGup10 I used Jev to replace a keyword filter, which never called a model at all. On 71 posts it kept 41 of the 45 on-topic ones, the keywords kept 10. So some of this demand is brand-new calls that don't come out of anyone's flagship bill.

[Kıvanç Yüksel @thejoyrises]

An IP kept topping our weekly security review as an anonymous /api/user caller. It was our own prod server.

Pricing page server-renders, finds no user, falls back to fetching /api/user at the public origin. The call leaves the box, crosses Cloudflare, comes back. ~376/week.

[Kıvanç Yüksel @thejoyrises]

@Hartdrawss The one I'd push back on is the checklist only ever getting longer. Aviation and surgical checklists are short on purpose, because a list nobody finishes protects nothing. Something has to age out, usually the failures the stack now makes impossible.

[Kıvanç Yüksel @thejoyrises]

Asked an image model for "preserve fine edges (hair), transparent PNG" and got the transparency checkerboard painted into the pixels.

Gemini's image models and gpt-image-2 can't emit alpha. What stuck was skipping generation and matting the original photo with BiRefNet.