Marcus Lewis (@mrcslws) header
Marcus Lewis (@mrcslws) profile picture

ML, programming, neuroscience. Previously @UMichCSE, @Microsoft, @Numenta. Now making Jupyter / Tensorboard / etc. easier to reach at https://xitter.gq/t.co/7ketvVRlch

San Francisco, CA
Joined February 2011
I feel so happy for anyone after they have their first baby or buy their first Model Y.
77
...my freshman year Michigan organic chemistry lab TA just became a US senator.
1
6
274
Hmm… decisions…
Can’t wait to pro and tell next week.
3
86
Brad from fireinabottle.net has been obviously hacked for like a month, spamming followers with messages. (Anyone who tags him and says he's hacked gets blocked.) reddit.com/r/SaturatedFat/co…
158
For me, a fun entry-point into this paper is just to observe: The transformer is a silly architecture for token generation. During inference, transformers operate with their hands tied behind their back. They do this because it makes training cheaper.
What if a foundation model could tell us how to modify its architecture to boost inference and reasoning instantaneously—without retraining? What if that tweak incurred near zero latency cost during generation and supported indefinite state tracking? arxiv.org/abs/2608.17981
6
1
1
18
5,111
The transformer builds up information, then essentially discards it, again and again. Each token starts fresh, not receiving the information built up by the deeper layers for previous tokens. Shallow layers can't access the information from previous tokens' deeper layers.
2
3
336
Finally convinced myself that non-distorting watermarking is real, a la Anthropic / Google SynthID. It's not just spin / marketing. This really is a "Monty Hall" like problem. (h/t @random_walker for that analogy.) Sharing here in case it helps anyone else.
1
2
262
On one hand, watermarking is "obviously" damaging to the output. The LLM does all this work to compute token probabilities... then you essentially perturb the probabilities? Of course that's bad! Every single token is using perturbed probabilities!
1
107
This tweet is for all the babies that got to experience The Odyssey in IMAX from inside the womb.
1
5
1,037
Tip: if you play piano with a toddler in your lap, you get to blame all the mistakes on the toddler.
1
235
New experience: going offline to have lunch with your toddler while your blog post is on the front page of Hacker News.
3
1
1
14
5,742
★ A native graphical shell for SSH I spent time rethinking how we should interact with servers, Raspberry Pi, robots, etc., then built it up piece by piece.
7
1
2
13
1,793
Thing #28 I didn’t know about toddlers: They choose a book and put it in your hand the same way you would choose a cartridge and put it into a Nintendo.
4
192
People talk about how having kids teaches them things about ML / AI, but for me it's more the opposite direction. ML gives me a vocabulary for kids. For example, when a grandparent gives him "noisy training data", I'm relaxed, because a little bit of "regularization" is good.
4
238
A fun part of having a kid is seeing random things from your past repurposed: - the Pavarotti YouTube video you saw in your college opera musicology class becomes his ritual teethbrushing music - the guitar capo you got in high school becomes his most interesting chew toy
1
1
117
Sigh. 11 years later, here's what happens when you type "Han Solo" into YouTube search.
Another day, another book spoiled by a Google product. Machine Learning = Intelligence - Tact
1
2
130
One good way to clean up vibe code: ask the agent to write Markdown files about certain cross-sections. For example, "messages-between-components.md" or "database-format.md" or "example-usage.md". Then, suggest improvements. You're like Plato, editing the shadows on the wall.
100
9 rings for the mortal men
2
121
A fun part of early parenthood is you keep noticing "second chances". You get another run through childhood. Another chance to be good at throwing a ball, to know the difference between maple and oak trees, to be able to do fun accents.
117
I made something. It's like a web view, but native. Rather than parsing HTML and running JavaScript, it fetches machine code binaries over the internet and runs them sandboxed. Here's the first app: a "Top" app that runs over SSH.
1
1
2
10
4,036
A fun use of LLMs: set a breakpoint in your optimized code, copy-paste the disassembly, ask the LLM to find and explain how it does [THING YOU'RE CURIOUS ABOUT]. (I just learned how the compiler handles big "switch" statements over lots of integer values)
2
1
167
Now whenever I create a set of message IDs, I use a contiguous block of numbers, because it generates super nice assembly. (The `switch` statement assembly simply subtracts the lowest message number, checks bounds, then uses that as an index into an address lookup table.)
68
Tip: The coding agents are pretty good at archaeology. Give them your old folders of rusty code, and they'll give you back something that works.
1
70