Burak Ozturk

Ph.D., EIT, ML enthusiast. UWO’25, BOUN’20, MEF’18. RTs/likes are NOT endorsement. He,him/il,lui

[Burak Ozturk @buroztk]

Cumulative advantage = Luck

[Burak Ozturk @buroztk]

@thsottiaux Ultraslow mode: Throttles token generation to ~5 tokens/sec so your usage lasts the whole week.

[Burak Ozturk @buroztk]

@thsottiaux Limits seem messed up with the Sign in with ChatGPT feature. Stream keeps returning: “The ChatGPT user has reached their Subscription Sharing usage limit,” even though the app is allowed to use 100% and I’m nowhere near the limit.

[Burak Ozturk @buroztk]

Gemini Flash is still ridiculously the best model at parsing complex documents for what it costs. The accuracy/cost is hard to beat.

[Burak Ozturk @buroztk]

Meta has made itself so fundamentally untrustworthy that I can’t imagine ever buying its hardware. Some reputations are hard to undo.

[Burak Ozturk @buroztk]

I turned my iPhone to grayscale to stop doomscrolling and it’s absurdly effective. It basically killed the joy of looking at my phone. Everything looks the same & nothing is screaming for my attention anymore.

[Burak Ozturk @buroztk]

@theo If I were to diagnose a LLM with ADHD, it would be Astra.

[Burak Ozturk @buroztk]

If users surge in Codex, it means time to switch to Claude or vice versa. They have finite compute, and as demand grows, they prolly start serving different quantized versions to keep up with it.

[Burak Ozturk @buroztk]

They expected this to turn into a massive campaign that help with IPO down the line or get investors on board to boost T$ valuation. What they got is a damaged reputation where no logical researcher will use their models for drafts or discussion anymore. Local models are the way.

[Burak Ozturk @buroztk]

I keep saying for years “intelligence” is a buzzword that OAI and Anthropic keep using. Models are extremely good at connecting the dots when you point them in the right direction, BUT by themselves, it’s highly unlikely they will find a “unique” solution to a given problem. +++

[Burak Ozturk @buroztk]

Ever wonder why Blender/Unreal work suddenly became THE benchmark for frontier models?

My timeline keeps yelling “token efficient” while it eats my limits like cookie monster.

It’s marketing. And people happily ride the hype train because FOMO works. 🤷‍♂️

[Burak Ozturk @buroztk]

Remarkable work 🙌

[Design Arena @DesignArena]

BREAKING: MiniMax H3 Max sets the new Pareto Frontier for video generation, nearly 50x faster than the base model.

This model is post-trained by @fal on @MiniMax_AI H3, and it's in a league of its own: no other Image to Video model on the arena delivers higher preference at a lower generation time.

Its Image-to-Video generation time is just 6.4 seconds, 18x faster than average, and its Text-to-Video generation time is just 4.7 seconds, 24x faster than the average.

Huge congratulations to the @fal team on this release!

[Burak Ozturk @buroztk]

@thsottiaux

[Burak Ozturk @buroztk]

Even Opus 5 (High Effort) getting noticeably dumber is a signal that Anthropic is serving quantized versions of these models to save compute while training the next iteration. My guess is we’re 2–3 weeks max from the next iteration.

[Burak Ozturk @buroztk]

@weswinder $4 trillion dollar company btw

[Burak Ozturk @buroztk]

EU policymakers are so dumb they’re forcing companies to invent ways to detect LLM gen. text. It’ll surely hurt output quality, and at best, all these detectors can tell you is “maybe”

These people need a mandtry statistics course before talking about things they don’t understnd

[Burak Ozturk @buroztk]

@fkadev Instead of running a speed test every few minutes, why not just read the RX/TX byte counters? No extra data usage, and you can sample much more frequently.

[Burak Ozturk @buroztk]

I finally watched @obsessionmovie. It’s a masterpiece given the budget. I was hooked from start to finish. Curry Barker is definitely someone I’ll be keeping an eye on from now on. 🙌

[Burak Ozturk @buroztk]

GLM-5.2 is a big win for open source. If they manage to get these models running on accessible consumer hardware, the companies that poured billions into compute are cooked.

Soon? I don’t think so. But I’d love to witness that.

[Burak Ozturk @buroztk]

First step: retire the current Copilot and rebuild it from scratch. I genuinely don’t know a single person who enjoys working with a tool that feels this dumbed down.

[Satya Nadella @satyanadella]

https://t.co/vLmiBKTtX3

[Burak Ozturk @buroztk]

We need more open source LLM models. The biggest hurdle is access to compute. I genuinely hope governments invest more in making that computing power available to non-profits under truly open licenses.