Lukaku inanılmaz. Peri masalı gibi oldu

Tuna Hocamız Yalnız Değildir!

The impact of AI in four charts

👏👏

Bu adamla kurulan sevgi bağı çok acayip bir seviyede gerçekten…

Hayatımda duyduğum en saçma tartışmalardan biri. Üstelik bakanlar falan dahil olmuş. Yazık bu ülkenin enerjisine

Instant classic

Everything will be different from today on! #networkMax has landed

Nike spent millions on “Breaking2,” an attempt to break the 2 hour marathon by Eliud Kipchoge. Kipchoge did it in Nikes in 2019, but it was with lazers and pacemakers. It didn’t count. Today, 2 men do it officially. Both in adidas. The hits keep coming.

1-3 Nisan 1953 tarihinde yapılan BLUE SEA tatbikatına katılan TCG I.İnönü ve TCG Dumlupınar denizaltılarımız tatbikat dönüşü Gölcük'e intikal seyrinde, önde TCG Dumlupınar arkasında TCG I.İnönü olduğu halde 4 Nisan 1953 tarihinde gece 00:01'de Çanakkale Boğazı'na girmişlerdi.

Karpathy accidentally shipped the org chart for every AI-augmented company in 2030. Three files. program.md is the human writing strategy in plain English. train.py is the agent executing, iterating, and shipping code. prepare.py is the locked evaluation layer that neither the human nor the agent can touch mid-run. That third file is the one worth studying. In most companies deploying AI agents today, the person who sets the goal also controls how success is measured. The marketing team picks the KPI, runs the campaign, and reports the results. The PM defines the metric, ships the feature, and presents the dashboard. The incentive to subtly shift the goalposts is built into the structure. Karpathy separated goal-setting from evaluation by making prepare.py immutable. The agent optimizes val_bpb. The agent cannot redefine val_bpb. The agent cannot swap in a friendlier dataset. The agent cannot adjust the tokenizer to make its numbers look better. It either improved on the locked metric or it gets reverted. No narrative. No context. No "well, if you look at it this way." That's why the results held. 700 experiments, 20 kept, and when Karpathy applied those 20 improvements to a model twice the size, every single one transferred. The gains were real because the agent had zero ability to make fake gains look real. Shopify's CEO ran the same architecture overnight. 37 experiments, 19% quality improvement, smaller model beating a larger one. The pattern transferred because the evaluation was trustworthy. Now scale the principle. A sales team where the AI agent writes outbound sequences, an independent system scores reply quality, and a human sets the targeting criteria. A product team where the agent ships variants, a locked analytics pipeline measures retention, and a PM writes the experiment brief. A recruiting team where the agent screens candidates, a calibrated rubric scores them, and a hiring manager defines the role. The separation Karpathy built into 630 lines of Python is the same separation every company will need when agents do the execution. Whoever controls the eval controls the outcome. Lock it down or the agent will find the shortest path to a number that means nothing.

Müthiş gol atmış Orkun. Geçen haftaki frikik golü de acayipti

Liverpool’da Van Dijk ve Salah dışında karakterli oyuncu kalmamış. Hele Arne Slot’um halleri ve konuşmaları tam bir şaklaban. Umarım hakettiği gibi yakında kovulur

By far the worst refreeing performance I’ve seen in a long time. Pawel Raczkowski did not officiate the game tonight, he was the puppet on the field

Güzel iş…

Max Dowman bu akşam PL tarihinde gol atan en genç futbolcu oldu. 2009lu lise öğrencisi. Çok özel bir yetenek

İbrahim Hacıosmanoğlu İSTİFA! #BJKvGS

I heard an explosion over Dubai (palm Jumeirah) ~13:35 local time