Martin JJ. Bucher (@mnbucher) header
Martin JJ. Bucher (@mnbucher) profile picture

phd student @stanford — prev @autodesk @eth

San Francisco, CA
Joined September 2010
Photo by Martin JJ. Bucher (@mnbucher)Photo by Martin JJ. Bucher (@mnbucher)Photo by Martin JJ. Bucher (@mnbucher)Photo by Martin JJ. Bucher (@mnbucher)Photo by Martin JJ. Bucher (@mnbucher)Photo by Martin JJ. Bucher (@mnbucher)Photo by Martin JJ. Bucher (@mnbucher)Photo by Martin JJ. Bucher (@mnbucher)Photo by Martin JJ. Bucher (@mnbucher)Photo by Martin JJ. Bucher (@mnbucher)
anything that can be streamlined into tokens will be overtaken by agents and swallowed by the bitter lesson—you either work at the frontier of intelligence, hum2hum vibes, or learn how to bake good bread. anything else seems a suboptimal move given the current status quo.
6
451
✨What if you could create and redesign scenes by just talking to them? **ReSpace: Text-Driven 3D Scene Synthesis and Editing with Preference Alignment** Add, remove, and swap objects simply via natural language, e.g., "add tufted dark gray sofa". (1/8)
2
7
1
38
6,781
At the heart of our method lies a Structured Scene Representation (SSR), encoding non-rectangular room boundaries and object semantics explicitly via text. Lightweight, interpretable, and editable. We then formulate scene synthesis and editing as next-token prediction. (2/8)
1
6
184
casually hitting the gym at stanford with donald knuth is actually WILD
2
140
so turns out MCP is essentially a reverse proxy that abstracts away all those tool use APIs from your system prompt ?
1
3
311
only at stanford can you take a class for what is *about* to happen, not what *has* happened.
4
162
plot confirmed.
Replying to @mikeknoop
What was the rush? Why come and effectively declare victory without doing some rigorous analysis? Isn't the whole point of this prize to be rigorous in its evals rather rush to meet an agenda or timeline of a particular participant? o3 is indeed very capable, everyone agrees, but aren't you genuinely curious, Mike, what makes this result good?
137
stable-diffusion-v1-5 is the real OG
2
94
babe wake up, new scaling laws just dropped
1
164
the european arrogance can only really be understood by leaving europe for > 6 months. what a wonderful enriching journey it has been so far. draghi's report confirms many of these thoughts and further exemplifies the cognitive bias that most europeans have while the truth is actually the opposite.
135
the use of the term 'sustainability' or 'sustainable' should be completely banned, especially in the US. the term has become meaningless over the past decade. 'saving' material or energy consumption by 18% will not get us anywhere and is a complete waste of attention.
3
145
once we set the agents free, there is no going back
106
base llama 3.1 405b before sft+dpo is MAD
1
1
150
rant mode before post-training is kinda crazy :D these are default params with temp=0.7 and top_p=0.9.
1
89
In our latest paper, we challenge the claim that *zero-shot* generative AI models outperform smaller LLMs for text classification — and empirically show that ‘small’ LLMs fine-tuned on a few hundred samples consistently beat larger, zero-shot models across diverse tasks. (1/7)
1
5
2
18
5,314
We fine-tune and evaluate a variety of language models for sentiment analysis, approval recognition, emotion detection, and political stance classification, using different text types such as news articles, tweets, and political speeches. (2/7)
1
141
this morning my spotify playlist started playing backwards and it felt like a scene from memento
1
65
very smooth
watch @waymo dodge a car pulling a blind unprotected turn
109
checking weather in sf involves a three-step process: estimated temperature curve, wind speed AND uv index (!!)
1
59
if it turns out that they polluted lmsys arena with some clever watermarking and an army of bot-based votings i wouldn‘t be surprised at all. maximizing elo rating for pr reasons
155
different beast
Apple just solved LLM hallucination. Incredible work.
210
signal-to-noise ratio way too low in the past 48 hours pls stop 🍓
89
sometimes i bike through san francisco in the morning and wonder if a fast takeoff would just happen over night and i wouldn‘t be surprised. any takeoff scenario will happen here first no doubts
88