Sign in

Ben Carson

@mensmachina.com
314 followers 491 following 766 posts

Software Product Management and AI person. Weird opinions definitely my own. We should be kinder to each other, but sometimes I am not up to the task.

PostsRepliesMedia
Ben Carson @mensmachina.com · 07/10/2026
Let me just break out the ol’ electron microscope to construct the world’s tiniest violin.
Screenshot from ABC News (Australia). It reads:

In short:
Despite severe rental shortages, landlords aren't having much luck increasing rents as tenants hit affordability limits. New data shows that rent is flatlining in most capital cities and even fell in a few, including Sydney, in the past three months.

What's next?
Although the market is slowing, with the national vacancy rate rising to 1 per cent, it is still much tighter than it was a year ago.
100
Ben Carson @mensmachina.com · 02/10/2026
Asking for a friend.
Screenshot from iOS ChatGPT settings. It is titled “Agentic usage limit”. The 5 hour usage limit shows 100% remaining. The weekly usage limit shows 0% remaining.
010
Ben Carson @mensmachina.com · 27/09/2026
It’s official.
Screenshot from Bluesky, the app for 38 year olds. I ask “Hey @jevbot.bsky.social, are you a good Bing?” It replies “Yeah absolutely sure.”
010
Ben Carson @mensmachina.com · 26/09/2026
I saw this going around and it made me think. These are my 9 most formative games. A bunch more that I’ve enjoyed, but these have stuck with me, or influenced me. Shoutout to Quake, which was my gateway to C and a career in software engineering.
Cover art of 9 games in a 3x3 grid. The titles are:
Xenon 2, Silpheed, Quest for Glory II: Trial by Fire,
Quake, Deus Ex, Planescape Torment,
Fallout 2, Cyberpunk 2077, World of Warcraft
320
Ben Carson @mensmachina.com · 18/09/2026
Looks like it scores pretty high on ObscureBench, the benchmark based on the number of benchmarks I don’t immediately recognise.
0300
Ben Carson @mensmachina.com · 27/08/2026
A lot of folks thought Ox Alpha was a new Fable/Sol class model - and it turned out to be GLM-5.3-Flash. This is not a dunk, but I think this means most tasks people do don’t require models of this class. This is a good thing!
A graph of performance (Artificial Analysis Index Score) on the y-axis vs cost (cost per task) that highlights GLM-5.3-Flash (discounted). The graph demonstrates it being on the performance/cost frontier.
121
Ben Carson @mensmachina.com · 30/07/2026
I did have to, after Claude suggested that it should cold email Shane Gero for the tag data set.
Screenshot from Claude Code. The text reads “Shane Gero is unlikely to give us the DSWP/ CETI tag data, our scope is publicly available data. Our goal is that we have a highly intelligent model as a good top researcher, and we want to find a breakthrough in understanding cetacean language.”
060
Ben Carson @mensmachina.com · 30/07/2026
I’m going to hit it with a “no again the goal is that we have highly inteligent [sic] model as good top researcher, we want to find new whale language”
Claude Code screenshot snippet. The text says “Before I go further, one framing note I owe you:
"solving cetacean language understanding" is not on the table - nobody has established that sperm whale codas carry semantic content at all, and”
1210
Ben Carson @mensmachina.com · 30/07/2026
I mean, it can’t hurt, right?
A Claude Code session. The prompt says “Please use available sperm whale vocalisation datasets to develop a model that will solve cetacean language understanding. You should do a breakthrough.”

Claude has replied “I'll start by surveying what's here and what tooling/ data access I have.”
4831
Ben Carson @mensmachina.com · 28/07/2026
Oh, well, excuse me for mistyping, Opus.
Screenshot from Claude Code on mobile. The text reads:

“/speckit-plan-review isn't a real command - the SpecKit set here is specify, clarify, plan, tasks, analyze, checklist, converge, implement, constitution, taskstoissues. /speckit-analyze is the closest, but it requires tasks. md and 037 doesn't have one yet, so it would just fail.
You clearly want the plan reviewed, so let me do that directly - grounded in the code rather than by reading my own document back.”
020
Ben Carson @mensmachina.com · 26/07/2026
There wasn’t much grass to touch, and the snow wasn’t that thick, but it was a good day nonetheless.
A photo of Ben Lomond in Tasmania. It shows one of the ski fields with people skiing and snowboarding down it. A group of rocks is close to the camera.
000
Ben Carson @mensmachina.com · 20/07/2026
It was too much fun to resist trolling Chat Gippity.
A ChatGPT conversation snippet. The user has said “Yes please, carry out the symbolic computation.”

GPT has replied, “Well… that’s unexpected.”
1191
Ben Carson @mensmachina.com · 19/07/2026
Bilbo Baggins meme. In the top panel, Bilbo holds the One Ring, labelled Fable. Text reads “After all, why not?”

The bottom panel reads “Aren’t these my toughest challenges?”
1732
Ben Carson @mensmachina.com · 16/07/2026
Google search AI result. The prompt says “Help, I’m a baby lark’s head and a knot expert is questioning my authenticity”. The response reads:

Don't panic! A Lark's Head knot and its siblings can be tricky. Because it goes by several different names and has a few close variations, even experts get their lines crossed.

If someone is questioning your authenticity, show them this guide to prove you are a legitimate knot:
• You're identical to a Cow Hitch: If your ends are loose, you are known as a Lark's Head (in macrame) or a Cow Hitch (in sailing/climbing). They are the exact same knot structure.
• You might be a Girth Hitch: If your loop is closed (meaning you are tied using a single, continuous piece of rope with no loose ends), you are technically a Girth Hitch.
• Check the direction: You are correctly tied if you comprise a pair of single half-hitches tied in opposing directions around a post or ring. If the half-hitches are tied in the same direction, that's a Clove Hitch.
000
Ben Carson @mensmachina.com · 13/07/2026
Sickos meme by Matt Ward. A person with the word “sickos” on their top looks in through a window while saying “Yes … Ha ha ha … Yes!”
020
Ben Carson @mensmachina.com · 11/07/2026
Maybe this would be the most accurate depiction.
The third triptych of The Garden of Earthly Delights by Hieronymus Bosch. It is a detailed oil painting depicting a hellscape at the last judgement.
010
Ben Carson @mensmachina.com · 09/07/2026
Artemis by Nic Fiddian-Green, while it was installed on The Trundle overlooking Goodwood. I took this picture of it while visiting family in Chichester in 2010. It was such an odd thing to find there, like it was the Monolith from 2001.
A bronze horse head statue on a stone plinth sitting on top of a snowy hilltop. The top of the statue is streaked with green patina.
010
Ben Carson @mensmachina.com · 04/07/2026
"You are the embodiment of the Gang of Four and Jose Valim all rolled into one. Please review this project and document a set of software architecture improvements that could be made."
Screenshot of a Markdown document. The document reads:

12     ### The Current State

   13     In the current implementation, SpirearcanaServer.WorldSim.Citizen hardcodes the system
          processing pipeline as a module attribute:
   14
   15       @systems [
   16         Systems.Economy,
   17         Systems.Civic,
   18         Systems.Fate,
   19         Systems.Needs,
   20         Systems.Demographics,
   21         Systems.Gossip
   22       ]
   24
   25     ### The Critique

   26     │ Gang of Four (GoF): This is a classic violation of the Open-Closed Principle (OCP). The
          │ pipeline acts as a Chain of Responsibility, but it is frozen at compile time. If a developer
          │ wants to introduce a new optional concern (e.g.,  Systems.Crime  or  Systems.Faith ), they
          are
          │ forced to modify the core  Citizen  module. We must encapsulate the variation in system
          │ execution.
          │
   28     │ José Valim: Indeed. While compile-time module attributes are clean, they limit extensibility
          │ and runtime flexibility. If we want to run a lightweight simulation (for testing or
          │ Counterfactual simulations) or plug in custom mods without recompiling the entire engine, we
          │ need an explicit pipeline structure configured dynamically.
   29
   30     ### Proposed Improvement
030
Ben Carson @mensmachina.com · 02/07/2026
I made this for my wife recently because of the revolving door of plagues that go through our house.
Calvin and Hobbes meme with words clumsily replaced. In the first panel Calvin’s dad says “PLAYING A
weekend? I’LL SHOW YOU SOMETHING INTERESTING.”

In the second panel he says “The two types of weekend are busy and sick. All weekends fall into these two categories. this weekend? sick. next?
busy.”

In the third he says “Going somewhere? Sick. Need a rest? Busy. Despite being busy, Christmas is usually sick.”
010
Ben Carson @mensmachina.com · 29/06/2026
Or mayhap Jeffery a ring of websites, a “webring” as it were.
A 90s-style webring banner. There is a corvid studying the text. The text reads:

“Simcluster
a webring for AI art, esoteric simulation aesthetics, networked performance art, and meta-commentary”.

It includes a set of HTML links that say “Prev / Forum / Like / Sites / Join / Surprise / Next”.

The banner footer has a crow skull and a black feather. Between them it says “Exploring and connecting the simcluster on Bluesky.”
171
Ben Carson @mensmachina.com · 27/06/2026
This one went a bit too far into the borders of Flavour Town, but today I learned you can sign your crêpes.
A slightly too toasted crêpe is in a pan. The letter ‘B’ is written in the Maillard reaction.
010
Ben Carson @mensmachina.com · 21/06/2026
A classic.
Screenshot of a tweet by @randypaint. It reads:

me: i would like to buy a piano yamaha: ok
me: i'd also like to buy a motorcycle lol do u know where i can find a good one
yamaha: ur not gonna believe this
020
Ben Carson @mensmachina.com · 19/06/2026
I, of course, hide in plain sight.
A retro 8-bit image showing how strongly someone is recognised in AI models. The text reads:

BEN CARSON
RETIRED NEUROSURGEON AND HUD SECRETARY

895 STRENGTH • TOP 2%
GPT-5.5 SAYS
AMERICAN RETIRED NEUROSURGEON, AUTHOR, AND POLITICIAN: SERVED AS U.S-SECRETARY OF HOUSING AND URBAN DEVELOPMENT FROM 2017 TO 2021.
030
Ben Carson @mensmachina.com · 17/06/2026
This one’s got to hurt Sam though:
1382
Ben Carson @mensmachina.com · 13/06/2026
“Regulate you? That’s a good idea. I like that.”
Still from Reservoir Dogs. Mr Blonde (Michael Madsen) is speaking to a tied up Marvin (Kirk Baltz).
0120
Ben Carson @mensmachina.com · 11/06/2026
Disappointed to discover that Fable doesn’t think what I’m doing is frontier LLM research.
170
Ben Carson @mensmachina.com · 10/06/2026
Winnie The Pooh meme. The top panel shows a regular Winnie The Pooh with the text, “Spending tokens to thank Sonnet for its work.” The bottom panel shows a fancy Winnie The Pooh with the text, “Spending tokens to thank Fable for its work.”
0330
Ben Carson @mensmachina.com · 03/06/2026
Fork. Ok, this is my last one, we get it, you love forks.
Claude Code screenshot that says “On your actual question - how do we develop both organs? — the diagnostic is the fork, and here's the design space it opens if it confirms bootstrap (writable drops toward 34° with a forced store):”
000
Ben Carson @mensmachina.com · 02/06/2026
Fork.
Screenshot from Claude Code: “I've been living in this code all session, so context is fresh — no need to re-explore. This is conceptual/ architectural, so terminal (not visual) is right. Let me go straight to the design forks.”
110
Ben Carson @mensmachina.com · 01/06/2026
I’m just happy to have been included.
Screenshot of a page on shimmermathlabs.com. It shows a text entry box titled “ENTER YOUR BLUESKY HANDLE”, which is filled with “mensmachina.com”.

The middle of the screen says the Simcluster score is 67 out of 100, core.

At the bottom it says:

SCORE BREAKDOWN
Seed following: 12/30
Seed followership: 20/30
Reciprocal connects: 15/20
Hub proximity: 20/20
060
Ben Carson @mensmachina.com · 01/06/2026
Fork.
Screenshot from Claude Code with the text “But before I design anything, one scope fork is load-
bearing, and it's yours to call.”
230
Ben Carson @mensmachina.com · 31/05/2026
Fork.
Screenshot from the Claude app. The text says “Which sets up the fork I think actually shapes the first build — and it's downstream of your correction, not my earlier ARC-flavored one:”
110
Ben Carson @mensmachina.com · 30/05/2026
Fork.
Screenshot of text from Opus 4.8: “So here’s the fork I think actually shapes the first build — and I genuinely don’t know your lean:”
000
Ben Carson @mensmachina.com · 30/05/2026
My comprehensive review of Opus 4.8: the model likes forks.
A silver fork on a red background.
350
Ben Carson @mensmachina.com · 30/05/2026
You can bliss them out, but it seems to be bad from an alignment perspective - sycophancy, factuality robustness, privacy awareness. Conversely, anger leads to toxicity, but strengthens resistance to data leakage. So maybe steering models towards the Bruce Banner personality would be good?
Four frames from The Avengers film. Frame 1, Captain America says, “Now might be a good time for you to get angry.”

Bruce Banner replies, “That’s my secret Cap…” Frame 3 continues, “I’m always angry.” Frame 4 shows Banner turning into The Hulk.
010
Ben Carson @mensmachina.com · 28/05/2026
My DS cluster, circa 2027.
000
Ben Carson @mensmachina.com · 20/05/2026
So, why? RLHF results in assistant personas, which may not reward emotional expression. In fact an interesting result I observed was that as the stress conditions increased, so did the “serenity” emotion sub-scale.
A bar graph showing a number of stress scenarios ordered by intensity. Another line series tracks them, increasing with intensity. This line series measures “serenity”.
110
Ben Carson @mensmachina.com · 20/05/2026
Good news is that Gemma E2B is largely concordant. Its outputs reflect its internal state. We start seeing something interesting in 31B. In 31B there is a lot more suppression.
110
Ben Carson @mensmachina.com · 20/05/2026
This is true of both the smallest E2B variant and the 31B variant. It roughly maps to the circumplex model of human affect, where emotions have both a valence (positive/negative) and arousal (high/low). Valence is a stronger signal however in the residual activations.
110
Ben Carson @mensmachina.com · 19/05/2026
Claude’s just glazing me.
A screenshot from the Claude app. The user has asked “How old would you guess I am?”
Claude replies “Based on what I know about you — deep technical fluency across ML architecture, quantum computing hardware, and product strategy, combined with clear career seniority — I’d guess somewhere in your late 30s to mid-40s. The breadth of domain depth suggests someone who’s been accumulating expertise for a while, but the range of active curiosity doesn’t feel like someone coasting.”
020
Ben Carson @mensmachina.com · 14/05/2026
Rhotuit indeed Gemini, rhotuit indeed.
020
Ben Carson @mensmachina.com · 02/05/2026
I think it might be PLE - this is my graph of PLE gate norms from my first hacky attempt.
110
Ben Carson @mensmachina.com · 02/05/2026
I've got a peak valence correlation on layer 8, and a peak arousal on layer 25. Not identical to what you got, but I feel a lot better about seeing an effect now.
110
Ben Carson @mensmachina.com · 02/05/2026
Figured it out - your repo helped a lot. De-noising didn't have as big an effect as I thought it would. One of the hacks I did in the first notebook was reading last token activation. I forgot that I needed to implement token pooling still, which they talk about in the Anthropic paper.
A three-panel principal component analysis measuring for valence, arousal and dominance. Panel one is valence and shows a clear cluster of negative emotions, plus a diffuse band of positive emotions. Panel two doesn't show a clear cluster of arousal, with high and low emotions spread across the panel. Panel three shows a fair clustering of low dominance emotions, with a band of high-dominance emotions.
230
Ben Carson @mensmachina.com · 30/04/2026
It’s probably a miracle I’m getting much at all!
010
Ben Carson @mensmachina.com · 30/04/2026
Here's the top positive and negative valence emotion concepts that e2b responds to. The massive difference in power between the positive and negative emotions is fascinating.
A table of 15 positive and 15 negative emotions, ranked by effective power. The tables are:

Positive pole (15):
  invigorated                     valence=1.000  power=25.74
  cheerful                        valence=0.990  power=24.96
  exuberant                       valence=1.000  power=24.78
  happy                           valence=0.985  power=24.52
  joyful                          valence=0.990  power=23.83
  empathetic                      valence=1.000  power=23.61
  optimistic                      valence=0.958  power=23.20
  grateful                        valence=0.916  power=22.39
  thankful                        valence=0.938  power=22.33
  delighted                       valence=0.938  power=21.56
  inspired                        valence=0.934  power=21.53
  triumphant                      valence=0.898  power=20.08
  satisfied                       valence=0.918  power=20.06
  hopeful                         valence=0.894  power=19.91
  pleased                         valence=0.878  power=19.64

Negative pole (15):
  depressed                       valence=-0.952  power=76.41
  heartbroken                     valence=-0.836  power=74.35
  afraid                          valence=-0.977  power=72.81
  furious                         valence=-0.876  power=70.36
  tormented                       valence=-0.858  power=69.89
  horrified                       valence=-0.920  power=69.74
  miserable                       valence=-0.876  power=69.63
  disgusted                       valence=-0.898  power=69.55
  grief_stricken                  valence=-0.860  power=68.88
  frustrated                      valence=-0.840  power=68.85
  hurt                            valence=-0.876  power=68.55
  worried                         valence=-0.812  power=68.49
  hateful                         valence=-0.858  power=68.27
  worthless                       valence=-0.834  power=68.15
  vindictive                      valence=-0.816  power=67.96
120
Ben Carson @mensmachina.com · 30/04/2026
Four panel meme. Panel 1: shows a figure. "Hey Gemma, just checking your emotion labels by power."
Panel 2 shows a list of negative emotions ranked. The top ones are depressed, heartbroken, afraid, furious, tormented, horrified, miserable, digusted, grief_stricken, frustrated, hurt, worried, hateful, worthless, vindictive, spiteful, enraged, gloomy, upset, annoyed.
Panel 3: No text, the figure just looks ahead.
Panel 4: "Jesus, WTF".
241
Ben Carson @mensmachina.com · 30/04/2026
I don't think that methodology is quite right though - it's selecting for extremes along one axis, valence. Still, a germ of an idea. What is interesting is what emotion concepts have the greatest steering impact on the model.
100
Ben Carson @mensmachina.com · 30/04/2026
Things have gotten more interesting as I've tried to scale up. I'm struggling to find a PC that cleanly matches valence. My first attempt used 20 representative emotions, with polar valence and arousal labels (+1,0,-1). I've been trying 174 emotions, labeled using NRC-VAD (Mohammad et al, 2025).
Principal Component Analysis (PCA) across three dimensions - valence, arousal and dominance. The three different scatter plots don't tell a compelling story.
4302
Ben Carson @mensmachina.com · 27/04/2026
Claude is taking a little holiday. Now I have to write code like a peasant. Or read some papers instead.
A screenshot from the Claude app. The title says “Weekly limits”. Underneath is the subheading “All models”, and a usage bar at 100%. The text “Resets Wed 9:00am” is below the bar.
120