Nick Radcliffe
LightNews LightNews
njr0.bsky.social
Nick Radcliffe
@njr0.bsky.social
53 followers 43 following 180 posts
Sceptic
Posts Replies Media Videos
njr0.bsky.social
I spent an an intensive month pair-programming with Claude Code, fully expecting to hate it and for it to be a waste of time.

I did hate it, but it was not a waste of time. It worked and I changed my mind.

checkeagle.com/checklists/n...
Claude code at startup after /context command showing 32% of tokens used and 22.5% reserved for autocompactification.
November 14, 2025 at 12:35 PM
njr0.bsky.social
A review of Blue Machine, by @helenczerski.bsky.social.

njr.prose.sh/blue-machine

One of the best books I have ever read.™

#oceans #physics #ecology #biodiversity #sustainability
Cover of the UK Hardback edition of Blue Machine, by Helen Czerski. Feaures the oceans, as foreground, in blue, with the major currents shown in gold. The land is shown as background, in black. The projection is chosen to keep the oceans as the central focus, with the land split in various ways, as the ocean more often is, to allow the shape of the oceans to be more clearly seen.
April 26, 2025 at 6:04 PM
njr0.bsky.social
This is an interesting perspective. It's not important, but I tripped over variantstens in the copy on your site. (I presume it should be "variants—tens" or "variants, tens".)
Screenshot of paragraph features "varaintstens" (highlighted). Content is: In practice, the most recent versions of these systems depend not only on massive caches of text and images generated and curated by humans but also on human judgment and knowledge in other forms. In particular, the systems rely on reinforcement learning from human feedback (RLHF) or its variantstens of thousands of human employees provide ratings of model outputs. They also depend on prompt engineering—humans must use both their background knowledge and ingenuity to extract useful information from the models. Even the newest ‘chain of thought’ models regularly begin from dialogue with their human users.
March 16, 2025 at 9:19 AM
njr0.bsky.social
I’ve been trying Kagi on and off. It’s a paid web search engine, available at ‪kagi.com‬ and as a default search option in the Orion web browser. It’s good and has found a few things for me I couldn’t find with Mojeek or DDG. (Google hasn’t found anything for me for years.)

Well worth trying out.
Stats from my Kagi trial account. 52/100 web searches used. 0/50 AI interactions used.
February 1, 2025 at 9:18 AM
njr0.bsky.social
Good overview of current state of LLMs from @simonwillison.net:

simonwillison.net/2024/Dec/31/...

He's more positive than I am, but balanced.

As to his boxed observation: I'm prepared to be persuaded, and I don't think they are valueless. I think they're misunderstood, dangerous, and misapplied.
Last few paragraphs of Simon's post: If we want people with decision-making authority to make good decisions about how to apply these tools we first need to acknowledge that there ARE good applications, and then help explain how to put those into practice while avoiding the many unintiutive traps. (If you still don’t think there are any good applications at all I’m not sure why you made it to this point in the article!) I think telling people that this whole field is environmentally catastrophic plagiarism machines that constantly make things up is doing those people a disservice, no matter how much truth that represents. There is genuine value to be had here, but getting to that value is unintuitive and needs guidance. Those of us who understand this stuff have a duty to help everyone else figure it out.
January 3, 2025 at 9:36 AM
njr0.bsky.social
You will have heard by now that 2025 is a square year. So I thought I’d plot atmospheric carbon dioxide from 1BCE to now as a square heatmap. It’s not looking great.

Scalable version & PDF at stochasticsolutions.com/gallery/heat....

#climate-change #CO2 #sustainability #climate-crisis
Heatmap showing every year from 1BCE to 2024 in a 45x45 grid. Everything is blue, representing values close to 270 parts per million, until the early 1900s. The second top row, repesenting 1935 to 1979 shows gradually increasing levels. The top row, representing 1980 to 2024, starts green (representing about 320ppm) and ends up bright red in 2024, representing just over 420 parts per million. Every year from about 2000 is orange or red, representing carbon dioxide concentrations over 350ppm. No year before about 1990 is.
January 1, 2025 at 9:17 PM
njr0.bsky.social
@theguardian.com You might want to fix the name of the Italian PM here:
www.theguardian.com/commentisfre...

(I believe she is Giorgia Meloni, not Giulia.)
Screenshot from The Guardian. Refers to the Italiam PM as Guilia Meloni.
December 30, 2024 at 7:37 AM
njr0.bsky.social
Our long international nightmare is over!

Someone fixed CSS while I wasn't paying attention.

stochasticsolutions.com/they-fixed-c...

#css #centre #center #vertical #html
Two squares, each with horizontally and vertically centred text. The top square says "Hey diddle diddle". The bottom one says "The text's in the middle". Underneath these is the full web page, with the surprisingly simple CSS to achieve it two different ways, both featuring CSS's display:grid, one using place-content: center and the other using justify-content: center and align-content: center.
December 28, 2024 at 8:56 AM
njr0.bsky.social
I was looking for a map of the world scaled by population.

Of course, the brilliant @ourworldindata.org's brilliant @maxroser.bsky.social as one (from 2018). Would love to see this kept up to date and time-lapse.

assets.ourworldindata.org/uploads/2018...
Map of the World with country sizes scaled by population (2018). From Our World in Data. Country positions and shapes are preserved where possible. Hard to summarise in words, but inevitably low density countries, like Russia and Canada appear much smaller than their geographic size. Densely population countries such asIndia, Indonesia, Japan, Nigeria, D. R. Congo and even UK are shown much larger than on equal area maps (and common non-equal area projections).
December 25, 2024 at 8:53 AM
njr0.bsky.social
I'm known (slightly) for criticising Jupyter and other computational notebooks. So here's something more positive: Suggested best practices for safe notebook use.

Clickable version at: www.tdda.info/best-practic...

Printable PDF at: www.stochasticsolutions.com/pdf/nbp.pdf

#jupyter #notebook #tdda
Notebook Best Practices. (all text). HTML version at https://www.tdda.info/best-practices-for-notebook-users. PDF version at https://www.stochasticsolutions.com/pdf/nbp.pdf Text is: NBP1 □ Ensure that your Notebook runs correctly after completion □ Develop the Notebook □ Take a temporary copy of the Notebook □ Clear the Notebook □ Run and confirm the results match the temporary copy □ Fix (if this is not the case) □ Clear the new Notebook again □ Commit the new Notebook to version control □ Rerun (so the Notebook contains the results) □ Delete the temporary copy NBP2 □ Store the (cleared) Notebook in version control (see NBP1) NBP3 □ Parameterize inputs and outputs at the top of the Notebook □ e.g. Set and use variables such as INPATH and OUTPATH □ Write the most important outputs to file (if not already done) NBP4 □ Replace some individual cells or groups of cells with functions □ Move them into an importable file, import and use □ Prioritize potentially re-usable code and code requiring testing NBP5 □ Write some tests □ Create a reference (regression) test for the whole process □ Create unit tests for the individual functions NBP6 □ Consider restructuring/extracting the code as a standalone script NBP7 □ Allow the parameters to be set from the command line □ Alternatively,readfromaconfigurationfile(e.g..jsonor.toml) NBP8 □ Consider using safer alternatives like Marimo and Quarto
December 18, 2024 at 7:37 PM
njr0.bsky.social
Thinking about Dijkstra's quote (image), maybe rather than coding or hacking we should talk about bugging.

I bug
You bug
He bugs
She bugs
We bug
They bug
Copilot BUGS LIKE CRAZY.
Quote: If debugging is the process of removing software bugs, then programming must be the process of putting them in. — Edger W. Dijkstra'
December 16, 2024 at 2:18 PM
njr0.bsky.social
Spot the unfortunate typo:

#observationswindows
Not a great one for alt text, but two Observation Windows (a training one with a 12-month "pre-"-period and a 3-month outcome period) and a deployment one with only a (later) deployment period. Both, sadly, are labelled "Observation Widow".
December 13, 2024 at 6:22 PM
njr0.bsky.social
I wrote up some slightly more refined thoughts about that @ourworldindata.org graph I was commenting on the other day:

Post: Log Graphs and Grokkability.
www.tdda.info/log-graphs-a...

#data #visualization #TDDA
A graph showing the incomes in the table on a y-axis with a log scale and tick lines at $100, $200, $500, $1k $2k $5k and $10k. The seven countries are displayed on the x-axis as an orange dot for the 10th percentile income and a blue dot for the 90th percentile income. There is a vertical line between these tow number for each country, annotated with ratio of the two incomes. Top: the previous graph redrawn in much the same manner but using a linear scale for the $y$-axis. Bottom: the ratios from the original graph plotted in a similar style, but with the poor group in each country at value 1 (1 times), and the rich group at whatever the income ratio is for that country.
December 11, 2024 at 10:35 PM
njr0.bsky.social
As the book on TDDA comes together, I am constantly surprised at the books and authors that end up in the bibliography. When I got up this morning, I did not expect to cite Robert Heinlein; yet this was the day it happened.
Extract from (in-progress) book: In his novel Stranger in a Strange Land, Robert Heinlein (1961) introduced the word grok. The definition that seems to me most closely to match its usage is: grok (transivitive verb) 1.To understand profoundly through intuition or empathy. grok (verb) 1. To have or to have acquired an intuitive understanding of; to know (something) without having to think (such as knowing the number of objects in a collection without needing to count them: see subitize). 2. To fully and completely understand something in all its details and intricacies. 3. To get the meaning of some-thing. — The American Heritage Dictionary of the English Language 2022 (5th ed.) Cover of the original 1961 hardback edition of Robert A. Heinlein's Strange in a Strange Land. Image is monochrome green image of the original hardcover edition, showing Rodin's sculpture, Fallen Caryatid Carrying her Stone.
December 9, 2024 at 5:10 PM
njr0.bsky.social
This is really interesting and important data, and an interesting visualization—subtler than it first appears.I've redrawn it using linear scales, but resorted to two graphs.

The use of a log scale is clever and enables them to show both the multiplier and the relative size together. But... 1/n
Two graphs showing the same data from the Luxembourgh Income Study as shown in the original post from Our World in Data, but on linear scales. The data compares the income of the 10th percentile and 90th percentile of after-tax income in seven countries The top graph runs from $0 to $8000 and shows income differentials for South Africa of 22x (c. $2470 to c.$110), Brazil 8.4x (c. $1650 vs c. $200), China 7.8x (c. $1900 vs. c. $250) Uraguy 5.6x (c.$2200 vs. c. $400) UK 3.8x (c. $4100 vs. c. $1100) UK 6x (c. $6830 vs. c. $1170) and Norway (c. $5120 vs. $1688). The bottom graph shows the multipliers on a linear scale, aligned with the top, with each country shown as a vertical line from 1x to its country multiplier.
December 7, 2024 at 4:11 PM
njr0.bsky.social
ffs (Fix filename spaces etc.) prototype.
Screenshot of ffs utility. Shows ls command with files with spaces, ampersands etc. in the filenames. Run ffs. Suggests mv command but a further ls shows it has not executed them. Run ffs -f and this forces the renamings. Final ls shows cleaned-up filenames.
November 23, 2024 at 7:34 PM