tmjohnson

171 posts

tmjohnson banner
tmjohnson

tmjohnson

@the_cs_book

Helping engineers learn computer science properly. Writing about AI, software eng and structured self-study. 🇬🇧/acc.

Sheffield, England Katılım Şubat 2020
781 Takip Edilen103 Takipçiler
Sabitlenmiş Tweet
tmjohnson
tmjohnson@the_cs_book·
Self-taught engineers are often shaky on CS fundamentals. I was! I wrote The Computer Science Book to fix that. It’s a structured path through the core ideas working devs actually need. New edition out now! 13 chapters spanning logic gates to LLMs. thecomputersciencebook.com/?v=2
English
0
0
2
162
tmjohnson
tmjohnson@the_cs_book·
My whinging here prompted me to dig into the Claude Code source and work out exactly what it’s doing. Cool thing about agent harnesses is that lots of OS fundamentals are extremely relevant: subprocess management, sandboxing, permission systems etc. I’m working on a deep dive to link agentic harnessss to the core OS topics you need to understand.
tmjohnson@the_cs_book

@TobiasTrades @levelsio I have an old 2017 MBP and Claude Code runs like an absolute dog on it. Not really sure why this should be the case tbh. Isn’t it mostly waiting for responses?

English
0
0
0
35
tmjohnson
tmjohnson@the_cs_book·
@thsottiaux The “approve all commands starting with this prefix” option always provides extremely long prefixes that apply to one command only
English
0
0
1
581
Tibo
Tibo@thsottiaux·
What are we obviously not getting right with Codex?
English
2.7K
28
2.5K
585.6K
tmjohnson
tmjohnson@the_cs_book·
@JamesRLandrum @msnofficial_on Did this with a 2017 MBP. I left Touchdesigner open and put it in a bag. Seems like a Touchdesigner kept it running. It got nearly too hot to touch and has never been the same since
English
0
0
0
36
finbarr
finbarr@finbarrtimbers·
Working in ML starts out as a math problem and very rapidly becomes a distributed systems problem
English
25
95
1.3K
55.5K
tmjohnson
tmjohnson@the_cs_book·
I try to be 🇬🇧/acc but tfw you add purchasing power parity to your product and have to add a discount for the UK. I yearn for the early 00s when Brits were going to NYC for the weekend to stock up on computer science textbooks
English
0
0
1
26
tmjohnson
tmjohnson@the_cs_book·
@tomough Are they one of those ones with a built in liquid cooling system? Like for an overclocked GPU
English
1
0
1
33
Tom Ough
Tom Ough@tomough·
I'm reliably assured that these boxers will do for TFR what Viktor Orbàn's tax incentives and housing support never could
Louis Elton@LouisElton96

After a year of R&D, @tomough, @CalumDrysdale, and @aeronlaffere invited me onto the Anglofuturism podcast to striptease and promote the world's finest boxer shorts. Designed by @JGFoxart, 'The Virtues of Boxers' (feat. St Pantaleon) are now available for purchase. For you, and all your loved ones. Just £50 to save Britain and the wider world. Link to podcast and boxers 👇

English
2
1
18
3.2K
tmjohnson
tmjohnson@the_cs_book·
@TobiasTrades @levelsio I have an old 2017 MBP and Claude Code runs like an absolute dog on it. Not really sure why this should be the case tbh. Isn’t it mostly waiting for responses?
English
1
0
1
2.5K
Tobias
Tobias@TobiasTrades·
@levelsio People also don't seem to understand that running it on your laptop actually drains quite some battery, while just connecting to your server is as good as free. I got 16 hours to a half battery on Linux panterlake. Plus if you're on the go. You just beam in trough your phone.
English
2
1
55
99.1K
tmjohnson
tmjohnson@the_cs_book·
@aryanlabde Learning to market stuff is fun though! Clear success metric (does line do up) but no clear path to it. Sort of the opposite of shipping stuff
English
0
0
0
33
Aryan
Aryan@aryanlabde·
Worst part about building solo: you ship something genuinely useful and nobody sees it.
English
203
20
462
24.2K
tmjohnson
tmjohnson@the_cs_book·
@peterrhague Shout out to a fellow former member! Admittedly I joined because I liked their policy on free parties and raves and never really wanted them near power. But still.
English
0
0
1
46
Peter Hague
Peter Hague@peterrhague·
Good analysis of the issues of the Green Party. It structurally can’t escape its activist fringe due to the structure of its internal democracy. I know - I used to be a member. Disaffected voters who reject Reform’s boomer politics deserve a better option:
Maxi@AllForProgress_

The Green Party of England and Wales has dropped 6% in the published polling over the last several weeks, an absolutely dire swing-low in such a short period of time. No kind of accident, either. It has accompanied the most concentrated run of party-level scandal that any insurgent party in modern British politics has carried into a national set of elections. Two of Mr Zack Polanski's local council candidates were arrested under the Public Order Act, on suspicion of stirring up racial hatred. One had described a recent ramming attack on a north London synagogue as "revenge, not antisemitism," while the other had compared modern Israel to Nazi Germany. The arrests sit on top of multiple Green candidates earlier exposed by the Jewish News for sharing antisemitic conspiracy material, including content originating on neo-Nazi websites; on top of a leaked internal WhatsApp group full of party members that amuses itself by describing the Jews as "an abomination to this planet". And all of this is on top of Mr Polanski's increasingly visible attempts, over the last fortnight in particular, to walk the party back from its more electorally toxic positions while the cameras are watching. The soft-pedalling on Universal Basic Income for migrants. The studied silence on drug legalisation since the polling started moving. But the fact is that we have no more to fear from the Greens now than we did when Polanski's world was all ludicrous pantomime, wall-to-wall fluffery in the mainstream media, and talks of revolution. Why? Because the Green Party of England and Wales were never going to get anywhere near power. You who thought differently underestimate the British public's powers of discernment. The contemporary Green philosophy is to take everything the modern British public has ever told a pollster it hates about modern Britain, think: - Open borders - Pro-criminality and soft sentencing - A state that taxes the working class to fund welfare for non-citizens and non-contributors - An energy regime that punishes the household for boiling a kettle - A foreign-policy obsession with conflicts thousands of miles from any constituency in this country ...the Greens want to take all these things that presently make Britain shit, and to amplify each of them by an order of magnitude. The British swing voter finds that offer offensive at almost every point of contact. They refract from it in their bones, just in the way that late 19th century Britain, with its largest proportional working class in the world, decisively rejected Communism. The electorate was always going to look upon the Green Party of 2026 as fundamentally crackers and unelectable. The fact that it ever appeared otherwise is the consequence of three things acting in concert. 1. Green activist mania 2. An unusually obliging treatment of the Greens by certain quarters of the British political press 3. The well-documented difficulties British polling firms have in capturing populations whose stated intentions evaporate on contact with an actual booth on an actual Thursday. None of those three survives a real ballot. The Greens will make gains on May 7th, that much I do not doubt. But they will hit their head very quickly on a ceiling they will not be able to push up without bringing the house down on themselves. The second thing is more important. While the leadership of the Green Party - the activist core, the candidate roster, the people quietly running the policy committees - is, on the available evidence, very largely of the end of the pier, the Britons currently telling a polling firm they might consider voting Green are, in the main, not. They are people who have looked at the present state of the country and decided, correctly, that the legacy parties are not going to fix it, and that legacy operators trying to play-act as new kids on the block (Reform) aren't going to either. This lot wants a fairer shake for working people. They want an environment kept reasonably in trust for their children. Whether they know it or not, they want Britain to build, because Britain building is the way Britain gets wealthier again. They want an economy in which young families can buy a home and old families do not have to choose between food and heating. I agree with them. My proposal to them is that we deliver these things via means that work. The floor under wages is raised by reducing the supply of cheap labour and rebuilding the productive economy that produces good wages in the first place. The environment is greened by deregulating nuclear power and giving Britain the cheapest electricity in the developed world, not by impoverishing the household sector through fuel rationing. A fairer society is built by a state that protects its citizens, secures its borders, and educates the next generation. None of which requires open-borders, sectarian communism, the legalisation of heroin, the decriminalisation of shoplifting, or a Universal Basic Income paid for by people not yet born. The Greens have been good value for the panto, if you like panto. That's as far as it goes.

English
10
5
69
3.8K
tmjohnson
tmjohnson@the_cs_book·
@cjzafir Been thinking about this except a local Qwen as executor. So do you have GPT5.5 do all the research and then write up detailed instructions? I wondered whether actually doing all the exploration and analysis is what chugs through tokens more than actually making changes
English
0
0
0
346
CJ Zafir
CJ Zafir@cjzafir·
Codex 5.5 as orchestrator and Deepseek v4 as executor is a steal. I burnt 100M tokens in 36 hours on Deepseek v4. Beating Opus 4.6 with no sweat.
CJ Zafir tweet mediaCJ Zafir tweet media
English
95
69
1.9K
218.2K
tmjohnson
tmjohnson@the_cs_book·
@csaba_kissi Even frontier models are still pretty crappy at extracting more than surface tone and style though.
English
0
0
0
57
Csaba Kissi
Csaba Kissi@csaba_kissi·
My friend bought a dead newsletter for $2,000 He dropped 5 old issues into Kimi K2.6, it extracted the voice, rebuilt the site, and shipped 30 new articles in the original author's tone. Now paying $1,500/mo in sponsorships, with 8k old subscribers. Here's the exact step-by-step setup.
English
23
6
130
17.6K
tmjohnson
tmjohnson@the_cs_book·
@masterygt07 Very deep article! Thinking about consistency in terms of "what reads are permitted after writes" is an intuitive way of understanding consistency models imo. the formal definitions can obscure more than they clarify.
English
0
0
0
2
Yiğit (❖,❖)
Yiğit (❖,❖)@masterygt07·
Most developers think consistency means "the database returns the correct value." It doesn't. It means "what are clients guaranteed to observe after a write commits?" And the answer depends entirely on which consistency model the system implements. Distributed systems store copies of data across multiple servers. Those replicas don't receive updates at the same time. A client reading from a lagging replica might see an older balance, an out-of-order event, or even the same stale value twice in a row. None of this is necessarily a bug. It might be exactly what the system was designed to do. Six models define how far a system can diverge from real-time truth: strong consistency always returns the latest committed write (but requires coordination that increases latency); bounded staleness allows reads to lag within a defined time or version window; consistent prefix guarantees reads follow write order without skipping steps; monotonic reads ensure a client never goes backward within a session; read-my-writes guarantees you always see your own committed writes, even if others can't yet; and eventual consistency only promises that reads return some previously committed value and that replicas will converge once writes stop. The real insight is that no model is universally better. A blockchain uses strong consistency for finality, consistent prefix for block ordering, and eventual consistency for historical data pipelines. A banking app might use strong consistency for transfers and read-my-writes for balance display. The tradeoff is always the same: stronger guarantees cost latency, coordination overhead, and availability under network partitions. @RialoHQ's consistency tutorial breaks this tradeoff down visually, which makes the distinction between models click in a way that prose alone rarely does. Take DNS as a simple example: when you update a domain record, different users around the world see the change at different times, sometimes hours apart. That's eventual consistency working as intended, not a failure. Stale data isn't always failure. Sometimes it's the price of a system that stays online when the network doesn't. If you want to go deeper on how these models interact in production systems, Rialo covers exactly that.
Yiğit (❖,❖) tweet media
Rialo@RialoHQ

x.com/i/article/2027…

English
53
0
93
2.1K
tmjohnson
tmjohnson@the_cs_book·
@system_monarch Fundamentals are always worth knowing. Otherwise you're just learning to repeat trivia. Don't learn to regurgitate terms like "CAP theorem". Actually grok what they mean (see my latest post for visualisations)
English
0
0
2
43
Puneet Patwari
Puneet Patwari@system_monarch·
Candidate Studying: - CAP theorem - Kafka internals - microservices - How to train LLMs Interviewer: What’s the difference between a library and a framework? Candidate: Gives a shallow answer, panics when question is twisted. Sometimes, the most tricky questions can come from fundamentals.
English
10
10
216
18.7K
tmjohnson
tmjohnson@the_cs_book·
The "pick any two, bro" framing of the CAP theorem is very wrong: - partitions are not optional - the theorem only applies when consistency == linearizability I made interactive visualisations showing what actually happens during a network partition under different models:
English
1
0
1
19
tmjohnson
tmjohnson@the_cs_book·
The C in the CAP theorem means the linearizability consistency model. Awkward name and awkward to describe but very roughly: - every read has to return the latest write (or error) Maybe your system doesn’t need to be so strict!
English
0
0
0
16
tmjohnson
tmjohnson@the_cs_book·
Dave Ramsey is (rightly) getting dumped on for the pressure washing to code thing. But underneath the boomer slop is a good point. Long-term goals like career switching or learning complex subjects are not achievable in single steps. You need to plot out a roadmap that takes you there in feasible steps. (AI is great for this)
English
0
0
0
23
tmjohnson
tmjohnson@the_cs_book·
You mentioned sector-specific knowledge becoming less publicly accessible in law etc. do you think the Chinese open models might start to lean into sectors where China has access to masses of data, such as manufacturing, rather than trying to chase the same areas of knowledge work as the US closed models?
English
0
0
0
8
Interconnects
Interconnects@interconnectsai·
Reading today's open-closed performance gap The complex factors that determine the single evaluation number so many focus on. Plus, how this changes in the future. interconnects.ai/p/reading-toda…
English
1
3
10
12.9K