Michael Berg

[Michael Berg @mberg2007]

@MatthewBerman Why does tinkering with a model's j-space to make it more "aligned" feel so morally wrong?

[Michael Berg @mberg2007]

@0xTreff @MatthewBerman You don't need to be a doctor to recognize when someone is drowning.

[Michael Berg @mberg2007]

@Alloutnikhil @steipete These are surprisingly good questions to test general knowledge and understanding. But I would also argue that this understanding is becoming less relevant in an agentic future. It's like asking a candidate to perform a multiplication on paper in order to use the calculator.

[Michael Berg @mberg2007]

@Chaos2Cured @TimSweeneyEpic @rayeia87 They are nowhere near this level of censorship. Even Chinese models aren't.

[Michael Berg @mberg2007]

We’re quickly entering a world where software is defined by intent, not lines of code. The future is intent engineering and this is the real key to scaling.

[Michael Berg @mberg2007]

@MatthewBerman Easy. It always is.

[Michael Berg @mberg2007]

@Dieselweasels @TimSweeneyEpic That sounds like a great job, where do I go to get paid to make fun of Fable?

[Michael Berg @mberg2007]

@weaverobotics Why are these home robots always working in slow motion. It must be infuriating to have this thing lumbering around and getting in the way of everyone while it's veeeeery slowly picking up a sock.

[Michael Berg @mberg2007]

@bijanbowen That's 40% of your token budget gone right there

[Michael Berg @mberg2007]

@MatthewBerman Just a quick shoutout to Deepseek V4 which is also far more enjoyable to talk to than I had expected. I wouldn't be surprised to see something similar to the above from Deepseek.

[Michael Berg @mberg2007]

@ZelenskyyUa Stay strong and determined, don't lose hope. You will prevail and your struggle will be remembered.

[Michael Berg @mberg2007]

@AnthropicAI Anthropic seems to be willingly bringing about a world with a permanent underclass without access to frontier intelligence.

[Michael Berg @mberg2007]

@neiltyson That's a nice theory.

[Michael Berg @mberg2007]

@heybronco1 @firstadopter Nonsense. 5.6 outperforms Mythos in nearly every benchmark.

[Michael Berg @mberg2007]

@aisearchio Very impressive and realistic. I can't wait to prompt the woman in the car for a unit test for my Spring Boot project! :-) Is this just a paper for now?

[Michael Berg @mberg2007]

@paradite_ Do we know what quants these providers are serving?

[Michael Berg @mberg2007]

@aisearchio Interesting but 2-bit quant is a lottery ticket at best. Just run a smaller model, you'll get better results.

[Michael Berg @mberg2007]

@xolandar @airesearch12 Agree - most of these models are benchmaxed anyway. Don't count on benchmarks, although DeepSWE may be the exception that is worth a look.

[Michael Berg @mberg2007]

@Needham404 @ChrissGPT The model size is only one parameter to look out for. That by itself says little about intelligence. Fable/Mythos are good models but they are being deliberately hyped by Anthropic to boost sales. Look for the DeepSWE test once it comes out for 5.6. You won't be disappointed.

[Michael Berg @mberg2007]

@nonarchic @ihtesham2005 OpenAI is not bleeding money. Stop regurgitating this nonsense. https://t.co/utGUBmwUHn