Diomidis Spinellis

This account is inactive following TU Delft's decision to suspend action on X for at least three months from 24 March 2025 onward, which also applies to faculty accounts. Please follow me at bsky.app/profile/coolsw…, @CoolSWEng" target="_blank" rel="nofollow noopener">mastodon.acm.org/@CoolSWEng, or linkedin.com/in/dspinellis/.

This account is inactive following TU Delft's decision to suspend action on X for at least three months from 24 March 2025 onward, which also applies to faculty accounts. Please follow me at bsky.app/profile/coolsw…, @CoolSWEng" target="_blank" rel="nofollow noopener">mastodon.acm.org/@CoolSWEng, or linkedin.com/in/dspinellis/.

As we prepare for #icse2025 in Ottawa, the #icse2026 team is excited to announce that the call for papers for Cycle 1 is open! Deadlines: March 7 (abstracts) & March 14 (submissions). Next year's ICSE will be in Rio de Janeiro (April 12–18). Participate! conf.researchr.org/track/icse-202…

Other lessons: bring-in professional moderators for discussions, be careful about controversial code comments, create a foundation for closing-down the system, open source is about community.

adopt Apple's contact tracing API, host on GitHub, squash individual commits between releases (security & privacy). Also: open source at the day of release rather than from the beginning (reduce noise).

Pragmatic insights (with which the audience's majority also agreed) by Terence Eden from open sourcing UK's COVID tracing app at #FOSDEM: Used MIT license because other departments already used it and it was short and easy for lawyers and the public to understand,

When a group of disgruntled academic researchers switches its career to Hollywood…

Ιn my Java streams lecture I accompanied a unique words example illustrating their use with a presentation of D. McIlroy's critique of D. Knuth's code in CACM's “Programming Pearls column” and an Indiana Jones scene as a parable. dl.acm.org/doi/pdf/10.114… youtube.com/watch?v=kQKrmD…

"Code Reading" by @CoolSWEng is that rare book that actually concerns itself with teching people how to read large codebases. Every newcomer in an organisation will be confronted with multi-million codebases on day 1 that they have to grok in order to do their job. Strangeley - this is not something I was taught at Uni. I mean - my code writing teaching was poor. (how could it be better when the teachers were not developers themselves? they simply didn't know) But code reading - that was not even on the radar. It occured to me too late in my Uni education that - "no one can teach you what they themselves don't know". Wish a good samaritan spelled that to me earlier. 😂

I thought that ChatGPT caused the engagement with StackExchange sites to crash, but it seems that the fall started way before its 2022 launch. Why do you think this happened?

@garrethleee History repeats itself: big-endian vs little-endian representations at the level of LLM tokens rather than machine words.

However, it was then rumoured that the earlier Claude models used a modified version of this three-digit tokenization. Instead of going from left-to-right, it goes the opposite direction, right-to-left! 👁️To better visualize this, imagine writing a really long number with commas. The numbers between the commas end up being the tokens. Intuitively, this grouping make sense, since it is how we typically read and interpret numbers. [5/N]

🚀 With Meta's recent paper replacing tokenization in LLMs with patches 🩹, I figured that it's a great time to revisit how tokenization has evolved over the years using everyone's favourite medium - memes! Let's take a trip down memory lane! [1/N]

@LorenaABarba Good point! They may feed them to ChatGPT to summarize 😃. Seriously though, some asked me to explain the feedback, so obviously they had read it.

@CoolSWEng Nice. But… will they read them? 😬

My ai-repo-feedback scripts used the OpenAI API to prepare and send out twenty 18-page formative feedback reports to student teams working on a term-long programming assignment. They run in less than four hours at a cost of about $0.85. github.com/dspinellis/ai-…

Now there's no excuse for not writing a proper Unix manual page for the tool you've developed. Just prompt ChatGPT with its usage summary and ask for the man page troff source code. Edit for clarity, accuracy, and conciseness, and you're done!

github.com/dspinellis/ai-… it is real fun and helpful when you are forgetting stuff @DSpinellis @CoolSWEng 🙏

@kfountou My point is that the pressure to boost these metrics combined with the social media timeline selection algorithms have resulted in them often showing how much an account or post has gamed the system.

@CoolSWEng I guess this depends on how “good” is defined. Large number of followers certainly implies something depending on the context. Same goes for citation count.

The metrics of followers, shares, and impressions on social media are a perfect demonstration of Goodhart's Law, which states that “when a measure becomes a target, it ceases to be a good measure.” Nowadays none of these metrics tracks reliably worthy people or content.

@CoolSWEng github.com/dspinellis/ai-…

No, it's actually Friday. The low code quality in embedded devices never ceases to surprise me.

Today's @ft article “The cautionary tale of Goldman and Apple’s credit card” is a good example of a point I stress when I talk about Agile practices. They result in better products built more efficiently, but they aren't suitable for every domain. ft.com/content/ede9f1…