Mark Kretschmann (@mark_k) 's Twitter Profile
Mark Kretschmann

@mark_k

Follow me for news about AI, technology, and fitness. I'm a tech optimist and software engineer. e/acc

ID: 8721292

calendar_today07-09-2007 10:56:30

28,28K Tweet

6,6K Followers

441 Following

Mark Kretschmann (@mark_k) 's Twitter Profile Photo

I was there, man, deep in the bowels of the Google DeepMind, December 2025, the air thick with that sterile hum of servers and the faint tang of overworked engineers' coffee breath. They called it SIMA 2: a Gemini-powered beast dropped into virtual 3D hellscapes, tasked with

I was there, man, deep in the bowels of the <a href="/GoogleDeepMind/">Google DeepMind</a>, December 2025, the air thick with that sterile hum of servers and the faint tang of overworked engineers' coffee breath. They called it SIMA 2: a Gemini-powered beast dropped into virtual 3D hellscapes, tasked with
Mark Kretschmann (@mark_k) 's Twitter Profile Photo

Original ā€˜Neon Genesis Evangelion’ Studio Gainax Officially Shutters After 42 Years! Gainax, the legendary Japanese animation studio behind *Neon Genesis Evangelion* and founded in 1984, has officially closed after 42 years. Bankruptcy proceedings concluded on December 10, 2025,

Original ā€˜Neon Genesis Evangelion’ Studio Gainax Officially Shutters After 42 Years!

Gainax, the legendary Japanese animation studio behind *Neon Genesis Evangelion* and founded in 1984, has officially closed after 42 years. Bankruptcy proceedings concluded on December 10, 2025,
Mark Kretschmann (@mark_k) 's Twitter Profile Photo

Getting really excited about Nano Banana Flash. Pro is good, but imagine having something almost as good (4K!), but much cheaper and faster! It's likely we'll get it next week, together with Gemini 3 Flash. That's because Logan Kilpatrick posted and then deleted this: šŸ“·āš”ļø

Mark Kretschmann (@mark_k) 's Twitter Profile Photo

OpenAI’s flagship model, ChatGPT-5.2 Thinking, now ranks as the most censored AI on the Sansa benchmark. Not a good look for OpenAI. The ever-stricter censorship and ā€œsafety guardrailsā€ are actively pushing users away. More and more people are switching to Gemini and Grok. 😠

OpenAI’s flagship model, ChatGPT-5.2 Thinking, now ranks as the most censored AI on the Sansa benchmark.
Not a good look for <a href="/OpenAI/">OpenAI</a>. The ever-stricter censorship and ā€œsafety guardrailsā€ are actively pushing users away.

More and more people are switching to Gemini and Grok.
😠
Mark Kretschmann (@mark_k) 's Twitter Profile Photo

Gemini 3 Flash rendered this SVG (the important XBOX Controller benchmark). The model is expected to arrive next week and will then obliterate the competition. It's extremely strong, and yet cheap and speedy! You're in for a treat by Google AI šŸŽ

Gemini 3 Flash rendered this SVG (the important XBOX Controller benchmark).

The model is expected to arrive next week and will then obliterate the competition. It's extremely strong, and yet cheap and speedy!

You're in for a treat by <a href="/GoogleAI/">Google AI</a> šŸŽ
Mark Kretschmann (@mark_k) 's Twitter Profile Photo

Grok by @xAI can now control the self-driving mode (FSD) in a Tesla, and it can take instructions by voice! You can even tell it multiple destinations for your route, and switch it around afterward, without even touching your screen. This is next level! Watch the video:

Mark Kretschmann (@mark_k) 's Twitter Profile Photo

The job of a Software Engineer is quietly turning into a Product Engineer. LLMs have made writing code cheap and fast. The hard part now is knowing what to build and designing an experience people actually want. The engineers who matter most will not just implement features.

Mark Kretschmann (@mark_k) 's Twitter Profile Photo

Reposting is the lifeblood of X. Without repost, no post would go viral and most posts would never reach your eyes. This is how the algorithm has always worked, though it may change in the future when AI takes over.

Mark Kretschmann (@mark_k) 's Twitter Profile Photo

After careful consideration I'm now convinced that GPT-5.2 is *not* the completely new pre-trained model by OpenAI. Rather, it's another iteration of the GPT-5 model (which is an iteration of...), with further post-training. The new model ("Garlic") is yet to come, ETA January.

After careful consideration I'm now convinced that GPT-5.2 is *not* the completely new pre-trained model by <a href="/OpenAI/">OpenAI</a>. Rather, it's another iteration of the GPT-5 model (which is an iteration of...), with further post-training.

The new model ("Garlic") is yet to come, ETA January.
Mark Kretschmann (@mark_k) 's Twitter Profile Photo

GPT-5.2 xhigh by OpenAI scores 0% (zero) on CritPt. CritPt is a research-level physics reasoning benchmark designed to test whether an AI can do genuine, expert-grade theoretical physics reasoning. The highest score is 9.1% by Gemini 3 Pro.

GPT-5.2 xhigh by <a href="/OpenAI/">OpenAI</a> scores 0% (zero) on CritPt.

CritPt is a research-level physics reasoning benchmark designed to test whether an AI can do genuine, expert-grade theoretical physics reasoning.

The highest score is 9.1% by Gemini 3 Pro.
Mark Kretschmann (@mark_k) 's Twitter Profile Photo

Unsupervised FSD is coming very soon from Tesla! This means that the car is legally allowed to drive itself without a safety driver sitting in front. There is also no "emergency remote control". The car is completely autonomous! A technical masterpiece many years in the making!