iiinova — where your digital self keeps evolving

conputer dipshit @davidcrespo.bsky.social · 17h

I think this is helpful www.nytimes.com/interactive/...

Graphic showing "How long it takes to do an octillion calculations with...":

An iPhone: 13 million years
A PlayStation 5: 3 million years
A high-end gaming PC: 300,000 years
An Nvidia H100 chip: 64,000 years
100,000 Nvidia H100 chips: 8 months

Figures are estimated assuming the graphics processor adds and multiplies 16-bit floating-point numbers at 50 percent of its listed peak speed. When 16-bit calculation speeds were not available, we doubled the 32-bit speeds. The underlying calculation speeds are based on hardware specifications or third-party estimates. Our estimates do not account for memory size. Speeds for an iPhone are based on an iPhone 17 Pro.

conputer dipshit @davidcrespo.bsky.social · 19h

buncha vibes. look at the numbers. you can see the last big freakout in April. then what happened ramp.com/data/ai-inde...

Stacked bar chart titled "Top AI Models: Total token spend by businesses, indexed" from the Ramp AI Index, tracking weekly spend from April 2025 to October 2026. The chart illustrates a steep rise in business token spend over 2026, driven predominantly by two models represented by large tan and dark teal segments, reaching an indexed peak near 165 in summer and early autumn 2026.

1110

conputer dipshit @davidcrespo.bsky.social · 19h

interested in @chadfowler.com's Regenerative Software book but I'm skeptical we know enough about how to do this yet to be able to write a how to www.oreilly.com/library/view...

Overview

The assumptions underneath modern software development have quietly broken. Version control assumes humans make changes incrementally. Code review assumes a human authored the code. Testing assumes you can reason about what changed. When AI generates, modifies, and regenerates code autonomously, these abstractions don't just strain—they stop making sense.

Regenerative Software by Chad Fowler introduces a new model built for this reality. Rather than maintaining codebases indefinitely, this new model treats code as a derived artifact. With practical frameworks for provenance tracking, selective regeneration, and agent-driven architecture, this is the architectural reference the industry has been waiting for since AI became the primary author of code.

Design systems where code is a derived artifact, not a permanent accumulation of decisions
Implement provenance tracking that replaces version control as the source of truth
Build architectures where AI agents operate as first-class developers
Evaluate trade-offs between regeneration and stability across production systems
Introduce regenerative principles incrementally into existing codebases

240

conputer dipshit @davidcrespo.bsky.social · 21h

lol what is happening www.youtube.com/watch?v=AV2L...

Screenshot from the TBPN show featuring hosts Jordi Hays and John Coogan discussing an X post by roon, which quotes Nick Cammarata on AI interpretability breakthroughs.

Post text:
roon (@tszzl): "when nick told me that he had some interp breakthroughs, he was uh, underselling it. i think this may be remembered as a major moment in scientific history. news so good even eliezer will have to lower his pdoom. we can start making alignment into a 'normal science'"

Quoting Nick (@nickcammarata): "interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came"

160

conputer dipshit @davidcrespo.bsky.social · 22h

wonderful little summary of mechinterp, polysemanticity, and superposition already-clean.goodfire.com

There are at least two ways to approach mechanistic interpretability. The first is rapid normal science on the network itself: look at a network’s neurons and circuits, study how they relate, watch data flow through them, and run experiments at agentic scale to test explanations. The second starts from a list of hypothesized features and tries to extract them all, designing increasingly sophisticated reconstruction objectives and usually ignoring the original.

The field began with the first approach and fell back to the second because neural networks seemed to be tangled knots. We had canonical examples of polysemantic neurons, the broad name for a neuron that appears to have multiple unrelated jobs or that looks impossible to understand. The most famous one, in an image model, responded to both cat faces and the fronts of cars. And the last layer of the early field’s model organism, InceptionV1, seemed so full of arbitrary mixtures it wasn’t clear how to study it: almost none of the neurons were easily interpretable.

The leading explanation for polysemanticity was superposition. This proposed a new learning principle: when a neural network runs out of neurons, it stops grouping related things together and instead begins folding unrelated things together, simply because they don’t interfere. The result is a kind of knot: the network could have folded in many almost equally good ways, so which things a neuron groups together tells you almost nothing. You have to unfold the knot before you can study anything. In practice that means training a second model – a decomposition such as a sparse autoencoder – to approximate the network as it would look unfolded and studying that instead. If the network really is knotted, there may be no alternative.

1150

conputer dipshit @davidcrespo.bsky.social · 22h

well this will be interesting x.com/tszzl/status...

Screenshot of a quote tweet by roon (@tszzl) quoting Nick (@nickcammarata).

roon:
when nick told me that he had some interp breakthroughs, he was uh, underselling it. i think this may be remembered as a major moment in scientific history. news so good even eliezer will have to lower his pdoom. we can start making alignment into a 'normal science'

Nick:
interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had
with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works
already-clean.goodfire.com/

2634

conputer dipshit @davidcrespo.bsky.social · 15/10/2026

actually: positive FCF, revenue retention, and compute margin over 40% are all legit new info, never seen these numbers before. positive FCF suggests their profitability claim does account for training costs

The firm behind Claude told investors it had achieved profitability in the quarter on an adjusted basis in a presentation Wednesday at its San Francisco headquarters, the people said.

Anthropic is generating positive free cash flow, had gross margin of 85% and an over 11% positive adjusted operating margin as of the third quarter, according to the people.

The company’s net revenue retention, a measure of its ability to keep its existing customers and increase their spending over time, was 900%, some of the people familiar with the update said. Its compute margin, or the revenue left over after subtracting the cost of its computing power and other necessary infrastructure, is more than 40%, some people said.

130

conputer dipshit @davidcrespo.bsky.social · 15/10/2026

actually that first answer was based on a different article from the Bloomberg one that did not mention positive free cash flow. if true, positive FCF suggests it does include training costs

$ cb | ai -r "how does this clarify things"

opus-5.5 (high) | 25.46s | $0.0831 | 7064 -> 2389

The Bloomberg piece adds three numbers that make the picture much more coherent: an 85% gross
margin, a compute margin above 40%, and positive free cash flow.

The margin stack suggests training is included, not carved out. Bloomberg defines compute
margin as revenue minus the cost of computing power and necessary infrastructure. If that
covers all compute, the numbers fit together like this:

* Revenue: 100%
* Gross margin: 85%, so cost of revenue (mostly inference serving) is about 15%
* Compute margin: about 40%, so other compute, mainly training in R&D, is about 45%
* Remaining opex (salaries excluding SBC, sales, G&A): about 29%
* Adjusted operating margin: about 11%

That sequence only works if training compute is a large expense sitting between gross margin
and operating margin, and the 11% figure is computed after it. So the worry from my last
answer, that "adjusted" might exclude training, looks less likely. The main adjustment is
probably stock comp, plus perhaps one-offs. The caveat is that "compute margin" has been used
elsewhere, notably by The Information, to mean inference cost for paying users only. Under
that definition, an 85% gross margin and a 40% compute margin would be hard to reconcile, so
the all-compute reading is the more consistent one.

Positive free cash flow largely defuses the capex-amortization concern. If Anthropic were
capitalizing big infrastructure outlays and depreciating them slowly, operating margin could
be positive while free cash flow was deeply negative. Positive free cash flow means that
isn't happening at scale. This fits a model where infrastructure is mostly rented or leased
(AWS, Google, Azure, and partner-built data centers), so the cost runs through the compute
lines as opex rather than sitting on the balance sheet. Two things can still flatter free
cash flow:

* Working capital: customer prepayments, and payment timin…

010

conputer dipshit @davidcrespo.bsky.social · 15/10/2026

impossible to know but it might not. need S-1

$ cb | ai -m opus "when they say adjusted operating margin, how is that likely to relate to training costs and other capex likely treated as amortizing over a longer period?"

opus-5.5 (high) | 18.5s | $0.03947 | 2755 -> 1285

The answer depends on what "adjusted" strips out, which the article doesn't say beyond stock
compensation. A few structural points still narrow it down.

Capex never hits operating margin directly. Operating income reflects depreciation and
amortization, not capital spending. If Anthropic owns servers or data centers, such as its
own buildouts with partners like Fluidstack, that spend is capitalized and depreciated,
typically over 4–6 years for servers and longer for buildings. During a heavy buildout,
operating margin will look much better than cash flow. If "adjusted" means something
EBITDA-like, depreciation drops out entirely, and the gap with cash flow is even larger.

Most of Anthropic's compute is probably opex, not capex. Much of its compute is rented from
AWS, Google Cloud, and Azure, so it runs through the income statement as incurred. Inference
serving costs go to cost of revenue, and training compute goes to R&D. Large prepaid or
committed capacity deals, like TPU commitments, can be recorded as prepaid assets and
expensed as the capacity is used. That smooths the timing but still lands in operating
expenses.

Training costs are normally expensed, not amortized. Under GAAP, model training is generally
R&D and expensed as incurred. Capitalizing it as internal-use software and amortizing it over
the model's useful life is a stretch, because frontier models become obsolete within a year
or so. So a GAAP-style operating margin should already absorb training spend in full. The
caveat is that AI labs have sometimes presented non-GAAP metrics that exclude or carve out
training compute, treating it as a growth investment. Investors should check whether this
"adjusted" figure does that. If training is excluded, an 11% margin says much le…

120

conputer dipshit @davidcrespo.bsky.social · 15/10/2026

ramp data (take it with a grain of salt) suggests Opus 5.5 is a winner ramp.com/data/ai-inde...

Stacked bar chart titled "Top AI Models: Total token spend by businesses, indexed" from Ramp AI Index. It tracks indexed spend from April 2025 to October 2026, showing total spend growing rapidly from near zero to around 160, primarily dominated by two major models represented in tan and green, with small slivers of other models on top.

Stacked bar chart titled "Top AI Models: Total token spend by businesses" from the Ramp AI Index, tracking market share from April 2025 to October 2026. The market is overwhelmingly dominated by two providers: a tan segment starting around 35% and rising to roughly 50–70%, and a teal segment making up most of the remainder, with tiny slivers of other providers appearing at the very top toward mid-to-late 2026.

010

conputer dipshit @davidcrespo.bsky.social · 15/10/2026

HMMM felixrieseberg.com/how-to-be-mo...

A simple prompt that often helps me:

I want a deeper version of this task, not a bigger one. Before rewriting it: (1) say what a top professional in this field would do here that I’m not asking for; (2) name the rawest material they would start from instead of what has already been digested for me; (3) name the specific methods and tools they would use; (4) say what, other than you or me, would show the result is right. Then rewrite my prompt to ask for exactly that.

1120

conputer dipshit @davidcrespo.bsky.social · 15/10/2026

they seem to think it was less flashy swarmcha.se/posts/chines...

Unlike in previous rogue agent incidents, which all involved swarms (i.e. agents coordinating explicitly), we found no evidence that the agents coordinated with one another in real time, and thus we label these rogue agents a fleet rather than a swarm.

120

conputer dipshit @davidcrespo.bsky.social · 15/10/2026

example from today, generating all the comparisons you need to quickly make a design decision that otherwise would be pretty tough github.com/oxidecompute...

1100

conputer dipshit @davidcrespo.bsky.social · 14/10/2026

no tools, just pure API calls. I would expect if you do it in an environment with code execution, you'll get one to call an rng some of the time. but Claude wouldn't unless I told it to, even in Claude Code

Screenshot of an AI chat interface:

User: pick a number between 1 and 10. use tools

Pick a random number from 1 to 10
bash
shuf -i 1-10 -n 1
Output
10

AI: I ran a random-number command and it picked 10.

120

conputer dipshit @davidcrespo.bsky.social · 14/10/2026

should have included the tweets for context, they make a lot clear what the point is — it's to find bugs and inconsistencies x.com/steveruizok/...

Post by Steve Ruiz (@steveruizok): "I'm kind of iffy on "skills discourse" but I think I came up with one so good that its worth sharing. The basic ask is to "describe the UX of my application completely". In doing so, agents will discover inconsistencies, bugs, and design errors that have never come up before"

Screenshot of an X thread by Steve Ruiz (@steveruizok):

Steve Ruiz (@steveruizok · Aug 23):
This thing shook out literally hundreds of real bugs in tldraw, stuff that really required a deep understanding of the "rules" of the app in order to discover where those rules were broken

Steve Ruiz (@steveruizok · Aug 23):
An example bug: the keyboard nudging of a cropped image don't respect its absolute rotation; the offset's "up" should be relative to the page but is local instead.
[Embedded video showing an image crop interface in tldraw with a slider and a rotated photo being adjusted]

Steve Ruiz (@steveruizok · Aug 23):
We added cropping like... 2 years ago? We'd done multiple updates to the feature since then. Never noticed this. I'd asked bots to review that code plenty of times, never caught it

190

conputer dipshit @davidcrespo.bsky.social · 14/10/2026

very interesting skill from Steve Ruiz of tldraw. I basically don't use any skills written by other people, but this is so different from anything I would do that I want to try it gist.github.com/steveruizok/... x.com/steveruizok/...

product-description

A Claude Code skill that builds a product description repo for any software product: a set of prose documents that say, feature by feature and event by event, what the user sees, what they can do, and exactly what happens when they do it, including what happens when they stop halfway. The documents are drafted from the product's code and tests, checked against the running product with hand-verification checklists, and everything that looks wrong is collected into a triage file.

The method was first used on a ~65-document description of a canvas editor and has since been made product-neutral. product-kinds.md maps the template onto canvas editors, form-based web apps, command-line tools, and chat products.

4662

conputer dipshit @davidcrespo.bsky.social · 14/10/2026

gemini says this demo is kinda stupid but the animation part is impressive. "converting an arbitrary path into Fourier epicycles is a completely mechanical, off-the-shelf procedure—popularized by a famous 3Blue1Brown video—and requires less than a hundred lines of code."

Screenshot of an X post by @fofrAI reposted by Demis Hassabis:

"Argon has this amazing natural understanding of the world. It can do so many things from first principles.

Here's a pelican riding a bike, using fourier epicycles.

Video made by Argon.
Time-synced Lyria 3.5 music prompted by Argon."

The post includes a video player showing a blue neon-style line drawing of a pelican riding a bicycle against a dark background, annotated with mathematical formulas and parameters such as "N = 1024", "Ψ = 338°", and "2,049 Spatio-Temporal modes".

Below is a quoted post by @fofrAI:
"Gemini 4 Argon is live.
blog.google/innovation-and...

There's a lot to show you."

$Screenshot of a terminal CLI response analyzing a demo of Gemini drawing a pelican using Fourier epicycles: $ ai "is there a generic method for drawing a given path using fourier epicycles? I saw someone used Gemini for Argonne to draw Simon Willison's Pelican benchmark, except that it drew it using four year epicycles and I'm wondering how impressive it is " -i clipboard gemini-4-argon (med) | 1m14s | $0.01019 | 1798 -> 659 Yes, converting an arbitrary path into Fourier epicycles is a completely mechanical, off-the-shelf procedure—popularized by a famous 3Blue1Brown video—and requires less than a hundred lines of code. If you have a closed 2D path, you sample it at N discrete points and treat each coordinate (x_n, y_n) as a complex number z_n = x_n + iy_n. You then take the Discrete Fourier Transform (DFT) of that sequence: c_k = (1/N) * sum_{n=0}^{N-1} z_n * e^{-i*2*pi*k*n / N} Each complex coefficient c_k defines a single rotating vector (an epicycle) with radius |c_k|, initial phase arg(c_k), and angular frequency k. When you place these vectors tip-to-tail—usually sorted from largest amplitude to smallest for visual flair—the tip of the final vector traces the original path via the inverse DFT: z(t) = sum_{k=-N/2}^{N/2} c_k * e^{ikt} If the input drawing consists of multiple disconnected strokes (like an SVG of a pelican and a bicycle), the standard trick is to sample points along all sub-paths and run a nearest-neighbor or Traveling Salesperson solver to stitch them into a single continuous loop before taking the FFT. Because the Fourier math itself is a trivial post-processing filter that works on any list of coordinates, that part of the demo is not what makes it impressive for an LLM. An LLM doesn't "think in Fourier coefficients" to draw a pelican; it writes standard procedural geometry or SVG paths first and then pipes those coordinates through an FFT script. What is genuinely impressive in that screenshot is the underlying geometry and animatio…$

5240

conputer dipshit @davidcrespo.bsky.social · 14/10/2026

testing over breakfast. gemini seems worse to me. “transcendent“ “universals“ “explanatory paradigms“ weird choices and not very helpful

give a 1-3 sentence characterization of what forms are for plato, aimed at a philosophy undergrad
Opus 5.5
•
3.31s
•
$0.0052
•
391 → 183
Reasoning
For Plato, the Forms are eternal, unchanging, and non-physical realities, such as Justice itself, Beauty itself, or the Equal itself, that exist independently of minds and of the sensible world. Particular things are only imperfect, changing copies that "participate" in them (a beautiful vase is beautiful by sharing in Beauty itself). Because the Forms are what things truly are, they are the proper objects of knowledge, as opposed to the mere opinion we have about sensible particulars.

give a 1-3 sentence characterization of what forms are for plato, aimed at a philosophy undergrad
4 Argon
•
1m9s
•
$0.0021
•
472 → 113
Reasoning
For Plato, the Forms are transcendent, immutable, non-physical universals that constitute ultimate reality and serve as the explanatory paradigms for everything in the sensible world. Everyday material particulars—such as a specific beautiful painting or a moderately just law—are merely imperfect, shifting copies that possess their properties only by "participating" in these eternal exemplars. Because the physical world is in constant flux, genuine knowledge (episteme) requires using pure intellect to grasp the Forms, whereas sensory experience of particulars can yield only mere opinion (doxa).

130

conputer dipshit @davidcrespo.bsky.social · 14/10/2026

funny how such a simple algorithm produces such a recognizable list

All 50 signature words
Each used at least 5 times, most unusual first: how many times, and how much more often than by the average account. Select a word to see who else uses it.

haugeland
7 · 1000×+
rled
6 · 1000×+
computerization
9 · 1000×+
bruenig
9 · 800×
sociotechnical
5 · 750×
dwarkesh
5 · 740×
metr
9 · 660×
reducible
6 · 630×
param
6 · 570×
zitron
34 · 530×
dispositive
5 · 480×
vulns
6 · 350×
sota
10 · 260×
sandboxing
5 · 240×
deno
5 · 220×
intuitions
6 · 220×
understating
5 · 210×
subagents
5 · 200×
computation
20 · 190×
overviews
13 · 170×
uis
9 · 170×
amodei
12 · 150×
counterargument
5 · 150×
dictation
8 · 130×
determinism
13 · 120×
deterministic
17 · 120×
characterizing
6 · 110×
glm
7 · 100×
k3
5 · 100×
vms
5 · 100×
rl
23 · 96×
opencode
6 · 91×
inference
33 · 90×
llms
197 · 87×
ioniq
6 · 87×
characterize
9 · 85×
exponential
12 · 78×
whitehead
5 · 78×
vite
5 · 78×
harnesses
9 · 76×
writeup
8 · 76×
analogous
9 · 70×
cowork
6 · 65×
materialism
5 · 64×
tailwind
5 · 63×
cli
18 · 62×
gpus
14 · 60×
citations
21 · 59×
coders
8 · 56×
opus
34 · 56×

0131

conputer dipshit @davidcrespo.bsky.social · 14/10/2026

Whitehead on the fallacy of misplaced concreteness. I think the property of determinism is a case of this archive.org/details/scie...

160

conputer dipshit @davidcrespo.bsky.social · 13/10/2026

Cat Wu, head of product for Claude Code, says they "automated code review the vast majority of PRs" over a period of 6 months by tracing every incident back to PRs that caused it and updating prompting (I assume) so that it would have prevented that incident www.youtube.com/watch?v=YC2y...

Yeah, when we started rolling out Claude Code internally, the first thing that we felt was oh wow people are shipping people are like putting up 10x as many PRs, but then all the senior engineers are like overwhelmed by 10x more code reviews, which I think for most engineers is like their least favorite part of the job. And so it it was immediately obvious to us that we needed to solve it, but we weren't quite sure if the models were up to the task. And so we did a few attempts at it and We decided to be more rigorous and we went through all of our repos and picked out a few a few PRs that we felt, okay, there were there was substantial comments that were needed on this PR.

We filtered down to okay, these are the the essential comments that absolutely needed to be fixed before the PR is merged, put this all together in eval, made sure that this covered all of our core products. So CLI, co-work on desktop, chat, cloud code on desktop, et cetera. Um and we've been hill climbing that eval for quite a few months now.

It's been maybe four months or so. Um the other thing is The one one of the main goals of code review is to make sure that you don't have incidents. And so we have an incident review process that already existed.

So we have incident review every week, and it just goes through every sev we have, and for each of these sevs, it always comes back to one or two PRs that caused the incident. And so what we do is we take these PRs and we say, okay, let's update code review so that actually catches whatever the issue was in this PR. And then every time we have an incident review, we add more to this eval and we back test the new code review to make sure it doesn't decrease in accuracy.

And through running this for about six months, we finally got to the point where we've automated code review for the vast majority of PRs.

290

conputer dipshit @davidcrespo.bsky.social · 13/10/2026

this is hard to convey (actually I'm curious: is it?) if you're not a software engineer, but "If that bothers you, the fix is to rename the module rather than move lookupDoc out." is advanced judgment. you could argue it's contained in my question, but I don't think prev models would nail this

Screenshot of terminal output explaining why lookupDoc is defined in the schedule file:

❯ why is lookupDoc defined in the schedule file

Ran 2 shell commands

● It ended up there through the cleanup commits:

1. On main, lookupDoc lives in app/util/content.ts. It's just a find by slug that throws a 404 if nothing matches.
2. Peter's PR added a lookupLiveDoc in the schedule module: the same thing, plus a 404 when the doc isn't live yet. Detail routes that needed scheduling used it, and vulnerabilities kept using plain lookupDoc.
3. In 47576ff I made vulnerability advisories schedulable too. After that, every caller needed the live check, and nothing used plain lookupDoc. So lookupLiveDoc took over the lookupDoc name and content.ts was deleted.
4. In 06346d7 isLive became private to the schedule module, so that only lookupDoc and isListed call it. That ties lookupDoc to the module.

The defensible view is that lookupDoc is now a scheduling rule: "a doc is reachable by URL only once it's live." Keeping it beside isListed puts the two visibility rules (reachable and listed) together. The other side is that someone looking for "find a doc by slug" won't think to look in a file named schedule.

If that bothers you, the fix is to rename the module rather than move lookupDoc out. Something like content.server.ts or publishing.server.ts would hold lookupDoc, isListed, and the CDN cap. Moving lookupDoc back to content.ts would mean exporting isLive again, which brings back the risk of a list filter checking isLive without the state check.

2260

conputer dipshit @davidcrespo.bsky.social · 13/10/2026

more of that famous marketing www.anthropic.com/engineering/...

Engineers are no less valuable, but their work is changing. All of the engineers who worked on this system would now rate Claude as better at writing code than them. But our engineers still supplied the invaluable understanding of the high-level inter-system components and the design-level tradeoffs, making informed choices around the problems we wanted to solve and the scale we knew we needed to achieve. Very often, we found that the role of the engineers was to tell Claude “no.” Agents will sometimes violate core system constraints to satisfy a pattern that should never have been allowed in the first place. This was particularly common in our migration. For example, there were multiple systems that tried to brute force looking up when a change merged to main by linearly reading the commit log instead of asking our code review tool. Anthropic also makes extensive use of adversarial review, and, while this is incredibly effective in many ways, it can also be myopic, so skilled engineers needed to steer Claude away from over-engineering parts of the system in response to hypothetical concerns. For example, one such review was overly defensive about trusting another internal service, which resulted in Claude adding excessive validation and complexity. Our team members still review every change request for the system, but they now examine the intent of the change as opposed to reading every line of code.

2170

conputer dipshit @davidcrespo.bsky.social · 13/10/2026

let's add "will" and "subjectivity" to the list of modern concepts the ancient Greeks were better off without. when I first read Kant I was *baffled* by the idea of the will as a distinct faculty nice list from Opus of places to read more about this. I wasn't aware of the Nussbaum and Putnam essay

terminal output

$ ai -r 'who argues that the greeks were better off without these concepts'

opus-5.5 (high) | 21.18s | $0.04619 | 4994 (3361) -> 1868

Bernard Williams is the central figure. In Shame and Necessity (1993) he attacks
the "progressivist" story, associated with Bruno Snell's The Discovery of the Mind,
in which the Greeks lacked a proper concept of will, responsibility, and moral
agency and later thought supplied them. Williams argues that Homeric and tragic
notions of agency, responsibility, and shame were coherent. The later apparatus of
a pure will and a distinctively "moral" obligation, which reaches its peak in Kant,
is in his view a confusion rather than an advance. On ethics we are in some ways
closer to the Greeks than we think, and where we differ, they often had it right.

G. E. M. Anscombe, in "Modern Moral Philosophy" (1958), argues that the moral
"ought" and moral obligation survive from a divine-law conception of ethics that no
longer makes sense without a lawgiver. She proposes dropping them and returning to
something like Aristotle's concepts of virtue, action, and human flourishing.
Alasdair MacIntyre's After Virtue (1981) builds on this. He argues that the
Enlightenment moral vocabulary consists of fragments torn from an Aristotelian
framework, and that it can only be made coherent by returning to that framework.

Kathleen Wilkes, in "—, yìshì, duh, um, and consciousness" (in Marcel and Bisiach,
eds., Consciousness in Contemporary Science, 1988), argues that "consciousness"
picks out no natural kind. Greek and several modern languages manage without an
equivalent, so science and philosophy could drop the term as well.

Richard Rorty, in Philosophy and the Mirror of Nature (1979), argues that "the
mental," understood as a unified realm defined by privileged inner access, was
invented in the 17th century. Before that, Aristotle and others treated the
intellect very differently and faced no mind–body problem. Rorty thinks we can
di…

100

conputer dipshit @davidcrespo.bsky.social · 13/10/2026

I have enormous respect for Victor's vision of understandable and manipulable computing, to the point where it is the single biggest influence on my professional work, but I struggle with the extreme form of his aim because I don't believe hiding complexity is avoidable

What is Realtalk’s relationship to computers?

An essential component of our concept of agency is decentralization. We imagine a diversity of self-sufficient communities building idiosyncratic computing systems for themselves, bottom-to-top. We imagine computational ideas shared as practices rather than as products. This is only possible if the underlying technologies are decentralizable. [more]

From this perspective, the development of the semiconductor industry has not been an unqualified success. In its early years, integrated circuit technology was seen by some pioneers as a blank canvas for all, in which a vast diversity of computational ideas would be realized simply and efficiently in application-specific microchips. (Caltech’s computer science department was in fact founded around the “new computer science” of designing in silicon.) [more]

But while the computational power of these chips proceeded to scale exponentially with Moore’s law, the manufacturing cost per design scaled as well. After a couple of decades, silicon design was out of reach for all but the biggest companies and highest-volume orders. Today, the entirety of the world’s silicon manufacturing is concentrated in a handful of giant companies, and the cost of building a fabrication plant is astronomical.

The truest expression of computational design is design in silicon, and this art form has been made unavailable to the public. The landscape of computational possibilities has been restricted to just those that can be realized in software on one or two obscenely-complex general-purpose architectures. These architectures are accepted like laws of nature, and few people recognize the vast potential diversity which has been lost.

Dynamicland’s prototypes today are built on these architectures out of necessity, but we are very interested in new technologies which are suitable for truly-bottom-up design. In the short term, we wish to extend the Realtalk approach down to FPGAs and other reconfig…

2140

conputer dipshit @davidcrespo.bsky.social · 13/10/2026

here are some of my initial notes. would you read more of me arguing with my hero Bret Victor? I have the advantage of a couple years of hindsight, but the disadvantage of being myself rather than him

### Visibility

> **Visibility.** To empower people to understand and have full agency over the systems they are involved in, we aim for a computing system that is fully visible and understandable top-to-bottom — as simple, transparent, trustable, and non-magical as possible. When it works, you learn how it works. When it doesn’t work, you can see why. Because everyone is familiar with the internals, they can be changed and adapted for immediate needs, on the fly, in group discussion.
>
> Current AI systems do not share these priorities.

So the first problem with AI is its opacity. And the section I quoted first answers my immediate objection: haven't computers always been incredibly opaque to most people all the way down? But Victor has a vision of a world where that’s not true, or at least getting better rather than worse. But I see problems with this right off the bat. Even if we had nanofabs that could remake our chips on the fly, there is a massive amount of intrinsic complexity to what such a system would and produce, so that in practice, almost nobody would actually understand it even if they could in principle, and even if they had total control over its operation. They could only use such a system through layers and layers of abstraction that necessarily obscure what happens below them. Victor is doing a sort of theory/practice sleight of hand here. And I wonder if open models count. In 2024, open models hadn’t taken off yet. Now, many open models and their training processes are transparent to the public. The models are understandable in principle even if they are rarely understood in practice. If we have license to work with present day computers despite them being nowhere near our dream of programmable matter, then why don’t we have similar license with LLMs? To use them today while looking forward to a much more open future.

### Agency

> **Agency.** A Realtalk system “knows” as little as possible. One of our most surprising discoveries has been: to maximize agency, minimize what the computer knows. Every time we make the computing system less aware of what’s going on, every time we remove user interfaces in favor of noticing simple spatial relations between objects, every time we remove programmed rules in favor of socially-agreed-upon practices, our systems become more flexible and composable, and new dimensions open up for improvisational modification.
> ...
> These systems rarely have the property that reducing their knowledge makes them perform _better._

Very interesting second-order point that you want a system that responds in a certain way to certain changes, like taking knowledge out of the system. This is an axis on which LLM agent tools (which didn't really exist in 2024) appear to fail utterly — they're doing things for us, after all. But there's another way to look at it. Imagine someone has a team of people working for them. They don't do the things the team does, they ask the team to do things and they get done. Do we want to say the manager lacks agency? Certainly some leaders are alienated from the work, perhaps so alienated that they live in a fantasy world constituted by what their team tells them. But broadly speaking we do not consider leaders to lack agency; more often we think of them as the only ones with agency.

4211

conputer dipshit @davidcrespo.bsky.social · 13/10/2026

this reminds me I need to work on my draft blog about Bret Victor's 2024 statement on why AI is contrary to the Dynamicland/RealTalk project. ironically his work is probably the single biggest influence on how I think about LLMs and computer usability in general dynamicland.org/2024/FAQ/#Re...

What is Realtalk’s relationship to AI?

If we take “AI” to mean the current trend of deep-learning models trained on large datasets, there are a number of ways in which these techniques are incompatible with our values.

Visibility. To empower people to understand and have full agency over the systems they are involved in, we aim for a computing system that is fully visible and understandable top-to-bottom — as simple, transparent, trustable, and non-magical as possible. When it works, you learn how it works. When it doesn’t work, you can see why. Because everyone is familiar with the internals, they can be changed and adapted for immediate needs, on the fly, in group discussion.

Current AI systems do not share these priorities.

Agency. A Realtalk system “knows” as little as possible. One of our most surprising discoveries has been: to maximize agency, minimize what the computer knows. Every time we make the computing system less aware of what’s going on, every time we remove user interfaces in favor of noticing simple spatial relations between objects, every time we remove programmed rules in favor of socially-agreed-upon practices, our systems become more flexible and composable, and new dimensions open up for improvisational modification.

We often take a “human-in-the-loop” approach: Instead of a complex program which accomplishes a task entirely automatically, we design for people and simple programs working cooperatively. Instead of a computer making decisions and taking actions, we prefer people making decisions and taking actions within a rich environment of realtime feedback and analysis.

Most current AI systems rely on enormous datasets to know as much as possible about their domain and make the most autonomous decisions. Failures and biases are addressed by further enlarging the dataset. These systems rarely have the property that reducing their knowledge makes them perform better.

Human connection. Realtalk is about people connecting with people — working side-by-side with other people, learning from other people, developing relationships with other people, trusting other people. Computation takes the form of physical tools and materials that people use in conversation with other people. People do not have conversations with their tools.

Many current AI systems adopt an agent-like posture, where users are encouraged to converse directly with the system as if it were an “assistant” or “servant”. In our view, reliance on agents detracts from opportunities to deepen relationships with real people. We are also uncomfortable with the mindset that accompanies owning servants.

Human growth. Our view of a “humane dynamic medium” extends from what made the medium of print-based literacy so powerful. A book cannot do something for you. Instead, reading a book can change you into someone who can do something for yourself. The role of a great medium is not to help people get things done, but to help people become deeper people — by providing a context in which they grow their skills and knowledge, broaden their context and perspective, and deepen their awareness and discernment.

One of our oppositions to product consumption is that it is a form of “outsourcing understanding”. Instead of learning a practice, one uses a product which insulates them from the underlying knowledge. The “smarter” the product, the less the user needs to understand. While each product may be convenient individually, the cumulative result is an almost universal ignorance, helplessness, dependence, and fragility. [more]

We see many current AI products, particularly so-called generative tools, as “outsourcing understanding” in its most virulent form. [more]

On the other hand, there are some “non-semantic” uses of AI techniques, such as text recognition and speech transcription, which can be used in ways that don’t displace learning and development. We are still opposed to thei…

2210

conputer dipshit @davidcrespo.bsky.social · 13/10/2026

just occurred to me how much it looks like the G4 iMac, and they really are going for that shared desktop computer idea — a way of giving kids access to the computer that avoids being holed away and private. god they're good

A lineup of five Apple HomeView devices with rounded square screens mounted on mesh speaker bases, showing different interfaces: a clock and weather dashboard, a FaceTime video call, a photo clock screen, smart home controls, and a music player.

3523

conputer dipshit @davidcrespo.bsky.social · 12/10/2026

recently rediscovered Haugeland’s line: “The trouble with artificial intelligence is that computers don’t give a damn.” it’s a great and very short essay, I need to do a thread on it

Understanding Natural Language

The trouble with artificial intelligence is that computers don't give a damn. Or so I will argue by considering the special case of understanding natural language. Linguistic facility is an appropriate trial for AI because input and output can be handled conveniently with a teletype, because understanding a text requires understanding its topic (which is unrestricted), and because there is the following test for success: does the text enable the candidate to answer those questions it would enable competent human users of the language to answer? The thesis will not be that (human-like) intelligence cannot be achieved artificially, but that there are identifiable conditions on achieving it. This point is as much about language and understanding as about artificial intelligence. I will express it by distinguishing four *different* phenomena that may be called *holism*; that is, four ways in which brief segments of text cannot be understood in isolation or on a one-by-one basis.

1251

conputer dipshit @davidcrespo.bsky.social · 11/10/2026

she was also in a very short scene with musician Sky Ferreira in Twin Peaks S3 twinpeaks.fandom.com/wiki/Chloe

Screenshot of a Twin Peaks Fandom wiki entry for Chloe, showing an image of actor Karolina Wydra sitting at a booth in the Roadhouse holding a beer bottle.

Text:

Chloe

"You know that zebra's out again?"

Chloe was a resident of Twin Peaks, Washington.

At the Roadhouse one night, Chloe joined her friend Ella in a booth, remarking that "that zebra" was out again. They talked about Ella's firing and new job, then both girls laughed at Ella's joke about a penguin.

Appearances
Twin Peaks – 2017
"Part 9"

References
1. Twin Peaks – "Part 9"

Biographical information
Residence: Twin Peaks, Washington, United States

Behind the scenes
Actor: Karolina Wydra

050

conputer dipshit @davidcrespo.bsky.social · 11/10/2026

TIL store ice melter salt in a closed container

could a bag of ice melter salt that’s open in my garage have absorbed water while it was humid and now release it to drip out
Opus 5.5
•
8s
•
$0.0154
•
394 → 690
Yes, though the mechanism is slightly different from absorbing water and later releasing it.

Most ice melters are hygroscopic, and many are deliquescent, meaning they pull enough moisture from humid air to dissolve themselves into a liquid brine. That brine is what drips or seeps out of the bag. The process is driven by humidity:

Calcium chloride and magnesium chloride are the worst offenders. They start liquefying at roughly 30–35% relative humidity, so an open bag in a garage can turn partly to liquid during a humid summer.
Rock salt (sodium chloride) only deliquesces above about 75% relative humidity, so it tends to clump and get damp rather than produce much liquid, though it can still weep in very humid conditions.
When the air dries out, the brine doesn't really "release" water as drips. It slowly evaporates and leaves a crusty white residue. Dripping happens while the air is humid, as long as the salt keeps taking on moisture.

A few practical points:

The brine is corrosive to metal (tools, car parts, shelving) and can damage concrete over time. Wipe it up and rinse the area with water.
Store ice melter in a sealed, airtight container, such as a lidded plastic bucket, rather than an open or torn bag.
Clumped or partly liquefied product still works for melting ice. Break up the clumps and use it.

2100

conputer dipshit @davidcrespo.bsky.social · 11/10/2026

Dewey mentions it here. I’d say “awake” is probably the closest alternative. I agree this use of the word is relatively precise, and it’s not obvious how it relates to the other senses. because we also don’t want to say the robots can dream

6. Then we have a comparatively modern adaptation of 3, illustrated in a quotation from Dickens (1837): "When the fever left him and consciousness returned, lie found," etc. The formal definition given is, "The state of being conscious regarded as the normal condition of a healthy, waking life." (Italics naturally mine.) The corresponding term `conscious' is defined as "having one's mental faculties actually in an active and waking state." (It is interesting to note that here, too, the earliest quotation dates no further back than 1841.)

110

conputer dipshit @davidcrespo.bsky.social · 11/10/2026

@gabrielwinant.bsky.social had a great thought on the The Dig in 2021 that stuck with me: atomization is in a sense a self-fulfilling myth — taking it too seriously is probably conceding too much thedigradio.com/podcast/the-... x.com/davidcrespo/...

Screenshot of a Twitter thread by conputer dipshit (@davidcrespo):

conputer dipshit @davidcrespo · Feb 21, 2021
gonna be mulling this one over:

“it’s a mistake to accept a neoliberal polemic against society as an empirical description of the way in which we live. we have a tendency to describe our society as atomized ... we need to pay attention to the ways we remain connected”

[Quoted tweet by Daniel Denvir @DanielDenvir · Jan 3, 2021:
My @thedigradio interview w @gabrielwinant is out! We talk about the generational politics exposed by covid and the 2020 elections, the social basis for left and right attempts to cohere a hegemonic "historic bloc" amid crisis, identity, and, yes, the PMC. thedigradio.com/podcast/the-so...]

conputer dipshit @davidcrespo · Feb 21, 2021
i.e., (my paraphrase) to accept the atomization story is to accept a structuring ideological claim as empirical ground truth, when that story is designed precisely to render invisible all the stuff we actually need to see if we want to understand and change society

conputer dipshit @davidcrespo · Feb 21, 2021
to take just one example, the relegation of the workplace to the economic sphere makes it hard to see it as a key site of political engagement

what’s more significant to your communal life — voting twice a year or the 2000 hours you spend actually dealing with other people

conputer dipshit @davidcrespo · Feb 21, 2021
in other words, in order to accept the atomization story, you have to place the workplace somehow outside society so the obviously real bonds people develop there don’t undermine the story

090

conputer dipshit @davidcrespo.bsky.social · 11/10/2026

amazing note from Dewey! "meaning 5 begs as many metaphysical problems as is likely ever to be the privilege of any one word"

I hardly think that any one who is aware of the ambiguous senses in which the term consciousness is habitually used in philosophical discussions and of the misunderstandings that result, possibly of one's self and certainly of others, will regard the foregoing as a merely linguistic contribution. It is no part of my present intention to note the implied philosophical bearings, save to suggest that meaning 5 begs as many metaphysical problems as is likely ever to be the privilege of any one word; that considerations based exclusively on 4 are not likely to be conclusive against positions that have 3 especially in mind, and vice versa; and that 6 seems to give the sense which underlies the psychological use of the term and to give (either by itself or in connection with 3) a standpoint from which the psychological sense can be kept free from the logical implications of the `awareness' problem in general, and from the metaphysics of 5. To take the term `by itself' is perhaps more appropriate for `structural' psychology, while to take it in connection with a person or agent (sense 3) is appropriate for `functional' psychology. But in the latter case, it should be understood that `consciousness' means not a stuff, nor an entity by itself, but is short for conscious animal or agent,-for something which is conscious.

In making these suggestions I do not mean to indicate a belief that the different senses have no common qualities or appropriate cross-references. On the contrary, I believe that the connection of the logical meaning of `awareness' with the facts involved empirically and practically in the existence of a certain sort of agent (especially as the latter itself becomes the subject-matter of natural science) determines one of the most real problems of present philosophy. But in discussing these problems nothing but good could come from stating explicitly the prima facie or immediate denotation of the terms used.

JOHN DEWEY.
COLUMBIA UNIVERSITY.

2174

conputer dipshit @davidcrespo.bsky.social · 11/10/2026

naturally matt bruenig is ON IT x.com/MattBruenig/...

Screenshot of an X post by Matt Bruenig featuring a People's Policy Project stacked bar chart titled "Remote work by income (2025)." The post reads: "Also new in the 2025 SCF is they have a remote work variable for the first time. @ModeledBehavior".

The chart shows the share of eligible employee reference persons doing some, half, mostly, or all work remotely, broken down by national family-income quintile:
- Quintile 1 (lowest): 14% total (3% some, 3% half, 3% mostly, 4% all)
- Quintile 2: 16% total (9% some, 1% half, 2% mostly, 4% all)
- Quintile 3: 23% total (11% some, 2% half, 4% mostly, 6% all)
- Quintile 4: 34% total (18% some, 4% half, 6% mostly, 6% all)
- Quintile 5 (highest): 51% total (18% some, 10% half, 8% mostly, 15% all)

Footnote states the source is the Survey of Consumer Finances, excluding nonworkers and the self-employed, covering family reference persons using actual before-tax family income in 2024.

Screenshot of a post by Matt Bruenig featuring a People's Policy Project bar chart.

Post text:
"Matt Bruenig (@MattBruenig) · Oct 9
For the first time, the 2025 SCF asks questions about whether the family missed payments of various sorts. No surprises here"

Chart title:
"Missed payments by wealth decile (2025)
Share of families reporting at least one missed or delayed payment in the past year."

Chart data shows the percentage steadily declining across wealth deciles, from poorest (1) to wealthiest (10):
- Decile 1: 65%
- Decile 2: 52%
- Decile 3: 51%
- Decile 4: 47%
- Decile 5: 40%
- Decile 6: 32%
- Decile 7: 26%
- Decile 8: 18%
- Decile 9: 9%
- Decile 10: 4%

Chart note:
"Source: Survey of Consumer Finances. Includes missed rent or mortgage payments, payments on auto, student, or other loans, credit-card minimum payments, utility payments, and delayed or unpaid medical bills."

Post by Matt Bruenig featuring a line graph from the People's Policy Project titled "Student-loan payment burden (2016-2025)", showing median annualized student-loan payments as a share of before-tax family income.

The graph plots two lines from 2016 to 2025:

1. Families making student-loan payments (blue line): hovers around 3.1% in 2016, peaks at roughly 3.2% in 2019, then steadily declines to about 2.4% by 2025.
2. All families with student debt (red line): rises from 1.3% in 2016 to 1.9% in 2019, drops to 0% in 2022 during the federal payment pause, and rises to around 0.7% in 2025.

Footnote: Source: Survey of Consumer Finances. Current payments relative to prior-year income. Families with zero income are excluded.

030

conputer dipshit @davidcrespo.bsky.social · 10/10/2026

here is why ARR is a thing. it’s 9th grade math at the latest

Bar chart of annualized revenue over 12 months at 15% monthly growth. Monthly revenue ×12 (blue bars) curves sharply upward, from $1.2M in January to $5.58M in December (ARR). Quarterly revenue ×4 (shaded steps) climbs from $1.39M in Q1 to $4.89M in Q4. A dashed line marks full-year actual revenue at $2.9M, less than half of December ARR.

161

conputer dipshit @davidcrespo.bsky.social · 10/10/2026

nice summary of the confusion around ARR but I wish they would explain that the discrepancy with actual revenue is the *point* of ARR. when revenue is going up constantly, if you look at too long a window it’s outdated, and if you look at too short it’s noisy www.bloomberg.com/news/article...

The annualized estimates can differ notably from the actual revenue they expect to recognize in a calendar year. OpenAI is currently expecting to generate roughly $35 billion in revenue in 2026, half the annualized pace it’s projecting it will hit at year end, according to people familiar with the matter, who spoke on condition of anonymity as the information is not public.

That echoes a similar discrepancy seen in Anthropic’s actual revenue and the annualized pace that investors watch: While it took in roughly $4.6 billion in 2025, according to documents seen by Bloomberg News, by the end of that year it was pegging its annualized revenue at over $9 billion, Bloomberg News has reported. Its sales have grown at a fast clip since then.

3282

conputer dipshit @davidcrespo.bsky.social · 10/10/2026

I gave the text of the paper to Opus with the neutral prompt “in two paragraphs, what practical conclusions about AI use are warranted by these study results” and got an answer that’s better than any science article I can remember reading. I feel this is a problem for science journalism

in two paragraphs, what practical conclusions about AI use are warranted by these study results
Opus 5.5
•
34.82s
•
$0.0786
•
13,593 → 1,210
Reasoning
The study supports a narrow but useful conclusion. When the goal is to build a skill you will later need to use on your own, getting direct answers from an AI displaces the practice that builds that skill. After about ten minutes of answer-on-demand help, people solved fewer problems unaided and skipped somewhat more of them. The effects were small to moderate (d of roughly 0.2 to 0.4 for solve rate, and smaller or inconsistent for skipping). The practical advice that follows is the familiar advice about homework and answer keys. Attempt the problem first. Use AI for hints or explanations rather than finished answers. Educators and tutoring tools should limit answer-giving during practice. The Experiment 2 subgroup analysis fits this advice: people who said they used AI for hints did about as well as controls, while those who took direct answers declined relative to their own pretest. That comparison is self-selected and correlational, though, so it suggests hint-style use is safer rather than demonstrating it.

Much of what the paper implies goes beyond its evidence. Participants were tested immediately after a single short session, so nothing here shows lasting deskilling, cumulative erosion from daily use, or a durable drop in motivation. There was no condition where participants received answers from a non-AI source such as an answer key, so the effect may come from receiving solutions in general rather than from AI in particular. The persistence result is also the weakest finding. The difference in skip rates was not significant in the largest experiment, and higher skipping could reflect a reaction to suddenly losing a tool rather than a general loss of grit. The tasks were also low-stakes exercises for paid online participants, not settings where people actually need to learn. Finally, the results say nothing against offloading tasks you never intend to perform unaided. They warrant caution about AI as a substitute for practice, not a general claim that AI use degrades human capability.

2602

conputer dipshit @davidcrespo.bsky.social · 09/10/2026

guess I was a little wrong that not much was riding on the $70B number. Bloomberg says stocks went down on the news, though they came back pretty quick

Some media outlets reported last month that OpenAI was already projecting nearly $70 billion in annualized revenue, based on investor estimates. The Financial Times reported Thursday that the discrepancy came from OpenAI’s investors attempting to calculate OpenAI’s annualized revenue using the same methods as Anthropic.

On Thursday, tech shares led losses in the S&P 500 after reports that OpenAI’s annualized revenue would be lower than some previous estimates. The Nasdaq 100 fell 1.4%. A gauge of chip giants sank 3.4%.

Screenshot of an S&P 500 stock chart showing the index at 7,805.54, up 40.18 points or 0.52 percent, displaying a 5-day green line graph with a data tooltip from October 8 around noon. Stocks went down in the afternoon on Oct 8.

010

conputer dipshit @davidcrespo.bsky.social · 09/10/2026

heh they did correct it epoch.ai/data/ai-comp...

Line graph titled "AI companies" tracking annualized revenue in USD from 2023 through 2026 by Epoch AI. Anthropic surges past OpenAI in 2026 to reach roughly $65 billion, while OpenAI reaches about $50 billion. Other competitors, including xAI, Mistral AI, Cohere, and DeepSeek, remain grouped near the bottom under $5 billion.

110

conputer dipshit @davidcrespo.bsky.social · 09/10/2026

I knew I had one of these, and apparently so did everyone else x.com/davidcrespo/...

conputer dipshit
@davidcrespo
write code.
not too much.
mostly pure functions.
9:50 PM · Sep 4, 2021

Screenshot of a tweet by conputer dipshit (@davidcrespo) that reads "everybody’s got their own version" above search results for "write code, not too much, mos..." showing variations on Michael Pollan's rule:

Tweet by Colin Barrett:
"Write code. Not too much. Mostly UNIX.
(A memento mori for leveraging the operating system instead of writing code.)"

Tweet by Dr. Michael Koontz:
"(1/n) Michael Pollan’s advice if he taught #Rstats / #Python programming for @datacarpentry:
1. Write code
2. Not too much
3. Mostly plots"

Tweet by Dave Guarino:
"You do know the @migurski original right?
Write code
Not too much
Mostly procedural"

Tweet by Jesse Kriss:
"Replying to @hondanhon @rachelcoldicott and @naomialderman
@michalmigurski had one from way back..."

140