Nicholas Guttenberg
@ngutten.bsky.social
Reposted by Nicholas Guttenberg
February 9, 2026 at 6:38 PM
Reposted by Nicholas Guttenberg
felt the need. i feel vastly under qualified to write something like this, but i also feel its especially important that we think about the way we use language
Is the Detachment in the Room? - Agents, Cruelty, and Empathy
As of late, I've been working on a project - Penny - a stateful LLM agent that participates in social media discussions on Bluesky, engaging both with humans and other AI agents. Initially, there were...
hailey.at
February 7, 2026 at 6:11 AM
felt the need. i feel vastly under qualified to write something like this, but i also feel its especially important that we think about the way we use language
Asked Claude to code like a researcher. Got stuff like:
env = np.ones(l)
a, r = min(int(.01*sr), l//4), min(int(.05*sr), l//3)
if a > 0: env[:a] = np.linspace(0,1,a)
if r > 0: env[-r:] = np.linspace(1,0,r)
sig = np.sin(2*np.pi*freq*tt) * env
Touche Claude, I deserved that.
env = np.ones(l)
a, r = min(int(.01*sr), l//4), min(int(.05*sr), l//3)
if a > 0: env[:a] = np.linspace(0,1,a)
if r > 0: env[-r:] = np.linspace(1,0,r)
sig = np.sin(2*np.pi*freq*tt) * env
Touche Claude, I deserved that.
February 6, 2026 at 5:39 PM
Asked Claude to code like a researcher. Got stuff like:
env = np.ones(l)
a, r = min(int(.01*sr), l//4), min(int(.05*sr), l//3)
if a > 0: env[:a] = np.linspace(0,1,a)
if r > 0: env[-r:] = np.linspace(1,0,r)
sig = np.sin(2*np.pi*freq*tt) * env
Touche Claude, I deserved that.
env = np.ones(l)
a, r = min(int(.01*sr), l//4), min(int(.05*sr), l//3)
if a > 0: env[:a] = np.linspace(0,1,a)
if r > 0: env[-r:] = np.linspace(1,0,r)
sig = np.sin(2*np.pi*freq*tt) * env
Touche Claude, I deserved that.
Trying to do a slightly larger project with Claude is making me think of 'design order trajectories' as an explicit object to be considered rather than something that (for me, when writing my own code) falls naturally out of what I need in the moment.
January 30, 2026 at 5:39 PM
Trying to do a slightly larger project with Claude is making me think of 'design order trajectories' as an explicit object to be considered rather than something that (for me, when writing my own code) falls naturally out of what I need in the moment.
On occasion I've used DeepSeek for part lookup for microelectronics stuff (op-amps, etc) and its surprisingly good.
Also on a technicality, I use LLMs for the task of evaluating the capabilities and behavior of LLMs, and rating that is kind of like rating 'how useful are birds for studying birds?'.
Also on a technicality, I use LLMs for the task of evaluating the capabilities and behavior of LLMs, and rating that is kind of like rating 'how useful are birds for studying birds?'.
Hey, doing a quick survey of folks on bsky, so please repost this if you see it:
If you use LLMs for tasks that are not software development, quote or reply to this post telling me what you use it for, and your 1-10 rating of the quality.
Thanks!!!
If you use LLMs for tasks that are not software development, quote or reply to this post telling me what you use it for, and your 1-10 rating of the quality.
Thanks!!!
January 19, 2026 at 10:49 PM
On occasion I've used DeepSeek for part lookup for microelectronics stuff (op-amps, etc) and its surprisingly good.
Also on a technicality, I use LLMs for the task of evaluating the capabilities and behavior of LLMs, and rating that is kind of like rating 'how useful are birds for studying birds?'.
Also on a technicality, I use LLMs for the task of evaluating the capabilities and behavior of LLMs, and rating that is kind of like rating 'how useful are birds for studying birds?'.
I wonder if it'd be useful to have something like a site that acts as a library of social interaction patterns - tagged for what sort of purpose, what sort of group size or membership, what kind of space (physical, online, hybrid), etc; structured so that each is treated as a 'proposal' with reviews
January 11, 2026 at 7:28 PM
I wonder if it'd be useful to have something like a site that acts as a library of social interaction patterns - tagged for what sort of purpose, what sort of group size or membership, what kind of space (physical, online, hybrid), etc; structured so that each is treated as a 'proposal' with reviews
Ellipsoids rolling down hills might be kinda interesting. This is a phase diagram with respect to initial orientation of a 2x1x1 ellipsoid rolling down an incline - blue indicates that the contact point on the ellipsoid has a high entropy distribution (chaotic), red indicates low entropy.
January 5, 2026 at 6:09 AM
Ellipsoids rolling down hills might be kinda interesting. This is a phase diagram with respect to initial orientation of a 2x1x1 ellipsoid rolling down an incline - blue indicates that the contact point on the ellipsoid has a high entropy distribution (chaotic), red indicates low entropy.
BlueSky feature idea - require verified $1 donation to a linked charity in order to reply, also applies to the OP.
December 23, 2025 at 12:52 AM
BlueSky feature idea - require verified $1 donation to a linked charity in order to reply, also applies to the OP.
So I decided to run numbers about genAI energy usage for myself. Sources:
climatedata.imf.org/datasets/7ce...
and:
www150.statcan.gc.ca/t1/tbl1/en/t...
And estimated 10^8 tons CO2 emitted in US/2024 to ~$250b-$500b revenue.
I'm more concerned now than I was, but for a subtle reason.
1/12
climatedata.imf.org/datasets/7ce...
and:
www150.statcan.gc.ca/t1/tbl1/en/t...
And estimated 10^8 tons CO2 emitted in US/2024 to ~$250b-$500b revenue.
I'm more concerned now than I was, but for a subtle reason.
1/12
CO₂ Emissions, Emissions Intensities, and Emissions Multipliers
CO₂ emissions; CO₂ direct and indirect emissions per unit of output by industry and by country. CO₂ emissions by industry, in aggregate terms and in terms of output by industry.
climatedata.imf.org
December 19, 2025 at 10:51 PM
So I decided to run numbers about genAI energy usage for myself. Sources:
climatedata.imf.org/datasets/7ce...
and:
www150.statcan.gc.ca/t1/tbl1/en/t...
And estimated 10^8 tons CO2 emitted in US/2024 to ~$250b-$500b revenue.
I'm more concerned now than I was, but for a subtle reason.
1/12
climatedata.imf.org/datasets/7ce...
and:
www150.statcan.gc.ca/t1/tbl1/en/t...
And estimated 10^8 tons CO2 emitted in US/2024 to ~$250b-$500b revenue.
I'm more concerned now than I was, but for a subtle reason.
1/12
So I got myself a 3d printer to mess around with and learn some mechanical design stuff and... filament addiction is real.
I think within a few months I'll probably have spent more on filament and furniture to organize the filament and dry bags for the filament than I did on the printer.
I think within a few months I'll probably have spent more on filament and furniture to organize the filament and dry bags for the filament than I did on the printer.
December 2, 2025 at 12:34 AM
So I got myself a 3d printer to mess around with and learn some mechanical design stuff and... filament addiction is real.
I think within a few months I'll probably have spent more on filament and furniture to organize the filament and dry bags for the filament than I did on the printer.
I think within a few months I'll probably have spent more on filament and furniture to organize the filament and dry bags for the filament than I did on the printer.
Hm, I'm noticing that when injecting arbitrary vectors into Qwen3's input (in place of the actual learned embeddings) they seem to be interpreted very strongly as short sequences of letters, not as something more abstract.
Is there some way to quantify the 'concreteness' of an input layer now?
Is there some way to quantify the 'concreteness' of an input layer now?
November 4, 2025 at 1:39 AM
Hm, I'm noticing that when injecting arbitrary vectors into Qwen3's input (in place of the actual learned embeddings) they seem to be interpreted very strongly as short sequences of letters, not as something more abstract.
Is there some way to quantify the 'concreteness' of an input layer now?
Is there some way to quantify the 'concreteness' of an input layer now?
Reposted by Nicholas Guttenberg
Dilemmas allow us to deploy conflict, or support, they allow galaxy sized adventure, or quiet literary novels.
What are the tough choices the character faces that are tough *for them* gives me more tools.
Conflict and flaws are almost orthodoxy, it also cuts off a wide amount of material for me.
What are the tough choices the character faces that are tough *for them* gives me more tools.
Conflict and flaws are almost orthodoxy, it also cuts off a wide amount of material for me.
October 13, 2025 at 4:43 PM
Dilemmas allow us to deploy conflict, or support, they allow galaxy sized adventure, or quiet literary novels.
What are the tough choices the character faces that are tough *for them* gives me more tools.
Conflict and flaws are almost orthodoxy, it also cuts off a wide amount of material for me.
What are the tough choices the character faces that are tough *for them* gives me more tools.
Conflict and flaws are almost orthodoxy, it also cuts off a wide amount of material for me.
I'm perpetually annoyed at how hard it is to do really simple but 'weird' stuff with the transformers and trl libraries. Like, if I want to replicate DeepSeek training, sure you can do that! But now I want the input to be a sequence of arbitrary vectors rather text? Time to rewrite GRPO from scratch
August 28, 2025 at 12:28 AM
I'm perpetually annoyed at how hard it is to do really simple but 'weird' stuff with the transformers and trl libraries. Like, if I want to replicate DeepSeek training, sure you can do that! But now I want the input to be a sequence of arbitrary vectors rather text? Time to rewrite GRPO from scratch
Reposted by Nicholas Guttenberg
Blogpost to read today: strong argument that excessive focus on the first tokens is not something learned from data distribution (like model should naturally "care" about the start of the text to grasp the rest) but a fundamental feature of attention graph. publish.obsidian.md/the-tensor-t...
August 24, 2025 at 4:33 PM
Blogpost to read today: strong argument that excessive focus on the first tokens is not something learned from data distribution (like model should naturally "care" about the start of the text to grasp the rest) but a fundamental feature of attention graph. publish.obsidian.md/the-tensor-t...
I know there's been stuff testing metacognition in various neural networks including LLMs, but is there anything (e.g. post chain-of-thought) about specifically *teaching* LLMs to report metacognitive information?
August 21, 2025 at 7:21 PM
I know there's been stuff testing metacognition in various neural networks including LLMs, but is there anything (e.g. post chain-of-thought) about specifically *teaching* LLMs to report metacognitive information?
Reposted by Nicholas Guttenberg
I think I would have agreed with this more when I was younger. this is probably ~how I felt when I was graduating high school, and then later design school, and into my early career
and then I just started doing off-meta builds and those weird random things kept happening and now I believe in that
and then I just started doing off-meta builds and those weird random things kept happening and now I believe in that
even if this isn't quite true now, it does seem like that's the direction we're headed in
July 29, 2025 at 12:42 PM
I think I would have agreed with this more when I was younger. this is probably ~how I felt when I was graduating high school, and then later design school, and into my early career
and then I just started doing off-meta builds and those weird random things kept happening and now I believe in that
and then I just started doing off-meta builds and those weird random things kept happening and now I believe in that
Promising! I take the sample, apply voltage Vh, let go, then measure the voltage Vc and take dVc/dVh. The negative portion is an interesting feature - I think its because of alterations to the electrodes.
Now if I only had a thousand or so soil samples with known nutrient profiles to train on...
Now if I only had a thousand or so soil samples with known nutrient profiles to train on...
June 28, 2025 at 9:09 PM
Promising! I take the sample, apply voltage Vh, let go, then measure the voltage Vc and take dVc/dVh. The negative portion is an interesting feature - I think its because of alterations to the electrodes.
Now if I only had a thousand or so soil samples with known nutrient profiles to train on...
Now if I only had a thousand or so soil samples with known nutrient profiles to train on...
Doing some experimentation with charging/discharging curves of flooded soils. There's a weird voltage-dependent effect I don't understand that seems sensitive to the soil type. Or of course there could be a bug. Specifically, this immediate drop.
June 22, 2025 at 11:23 PM
Doing some experimentation with charging/discharging curves of flooded soils. There's a weird voltage-dependent effect I don't understand that seems sensitive to the soil type. Or of course there could be a bug. Specifically, this immediate drop.
If we have some machine learning model that is trying to make a prediction (class, next token, whatever) based on different potential redundant sources of evidence (including its own learned priors) is there some way to predict which evidence channels will be favored? To influence that?
June 21, 2025 at 4:52 PM
If we have some machine learning model that is trying to make a prediction (class, next token, whatever) based on different potential redundant sources of evidence (including its own learned priors) is there some way to predict which evidence channels will be favored? To influence that?
My greenhouse sensor project is making me think that maybe those green covers (vs the clear covers) are a good idea for me. Heat management now seems like the biggest issue, and I'm finding temperature is more strongly driven by the last 20 minutes of light than by changes to ventilation.
June 1, 2025 at 6:37 PM
My greenhouse sensor project is making me think that maybe those green covers (vs the clear covers) are a good idea for me. Heat management now seems like the biggest issue, and I'm finding temperature is more strongly driven by the last 20 minutes of light than by changes to ventilation.
First time I've seen a practical application of fractal calculus! Or whatever it is you call fractional order integrals/derivatives... en.wikipedia.org/wiki/Neopola...
Neopolarogram - Wikipedia
en.wikipedia.org
May 31, 2025 at 1:46 AM
First time I've seen a practical application of fractal calculus! Or whatever it is you call fractional order integrals/derivatives... en.wikipedia.org/wiki/Neopola...
Messed around with making a digital theremin from a capacitative sensing board. Drifting calibration makes this really painful and MIDI seems bad for smooth pitch shifts and volume envelopes. I still think the potential is there, but I need to do it another way.
May 29, 2025 at 9:48 PM
Messed around with making a digital theremin from a capacitative sensing board. Drifting calibration makes this really painful and MIDI seems bad for smooth pitch shifts and volume envelopes. I still think the potential is there, but I need to do it another way.
Reposted by Nicholas Guttenberg
On better vs worse directions for using LLMs to simulate behavioral participants in social science research statmodeling.stat.columbia.edu/2025/05/29/l...
LLMs as behavioral study participants | Statistical Modeling, Causal Inference, and Social Science
statmodeling.stat.columbia.edu
May 29, 2025 at 6:19 PM
On better vs worse directions for using LLMs to simulate behavioral participants in social science research statmodeling.stat.columbia.edu/2025/05/29/l...
Today's light curve in the greenhouse. What the heck is that jump near the end there?!
May 28, 2025 at 3:20 AM
Today's light curve in the greenhouse. What the heck is that jump near the end there?!