
How fast can you go through a JSON file from Python on CPU? You might be surprised at the answer... read my article to find out more.


Software performance expert. Ranked in the top 2% of scientists globally (Stanford/Elsevier 2025) and among GitHub's top 1000 developers. Father, husband.

How fast can you go through a JSON file from Python on CPU? You might be surprised at the answer... read my article to find out more.


A big security vulnerability in Java (CVE-2026-114938): a bug in OpenJDK’s AVX-512 Base64 decoder, present since 17.0.6, allows remote code execution. Base64 is a representation of arbitrary data (including images) into ASCII strings. It often ends with a padding character (’=’). Attached files in your emails are Base64 encoded. The bug has to do with our work. The decoder’s main loop follows Muła and Lemire, “Base64 encoding and decoding at almost the speed of a memory copy”: arxiv.org/abs/1910.05109 (with @pshufb). The bug is not in our algorithm or in its implementation. It is in the tail code. After the application of our algorithm on 64-byte blocks, the last few bytes are handled by a hand-coded assembly routine. They use masked stores, to write just the necessary data, as the fancy AVX-512 instructions allow, but they mess up the mask computation so that 64 bytes get written into a 50-byte array. Many mainstream systems, such as your browser, rely on an even fancier Base64 decoding algorithm… Fast Forgiving Base64 Decoding with SIMD Instructions ACM Transactions on Architecture and Code Optimization 24 (1), 2027 dl.acm.org/doi/10.1145/38592… It is in Bun, Node.js, and so forth, and under consideration in PHP. Much of my time with these high performance algorithms is spent testing. I spend serious money on servers to stress test the algorithms. (My GitHub sponsors help!) Mistakes are made, but we have better tools and we can do more analysis than ever, to write better and faster code.

I am getting suspicious. We are seeing surprisingly damaging software failures at a rapid rate. Could be pareidolia. Or not.
You may not be aware of what we're seeing in tech. The number of vulnerabilities has gone vertical thanks to new AI cyberweapons. Many colleagues believe the recent Microsoft/FAA/bank/NSA downtime is not accidental. You do need your best guys for this:

What should you think about universities making their lectures free on YouTube? About 20 years ago, pundits said that education would quickly move online, making the classroom obsolete. Several ventures were put together aiming to replace universities. Higher education would be cheap, maybe even free. We would school the entire world for nickels. It has mostly not happened. Online did happen. About half of all university students in the US take at least one online course during their studies. (Recent NCES data put this at roughly 55% in fall 2024.) But, if anything, the relative cost of schooling has gone up. Prices rose sharply for decades. And we now have AI, which is even more powerful than video. You can get your own professor to answer your questions. Isn’t that fantastic? Won’t that disrupt universities? Maybe not. About lectures. Students at Harvard mostly don’t attend classes. (Faculty reports and studies show high rates of skipping lectures; attendance often falls well below 50% by the end of term.) And you could always attend classes for free just by showing up to your local campus. Most lecturers do not object if you just show up in their classroom. Today, your AI is probably better than most lecturers at answering your questions. And it is always available. So, won’t this crush universities? Only if you think that universities are about education. But they are not. They are primarily about credentials. Here is the problem that they try to solve. You are young. Maybe 19. You can’t get a good job. What do you know? Now, of course, you can learn machine learning with your AI friend and videos, and more. But who is going to hire you? So you actually go to a good school and get a degree. Well. You are one step closer to getting an interview even if you learned little. Is this pessimistic? Not at all. I think that the future belongs to those who learn. In fact, AI just raised the bar. Just don’t confuse learning and a university education. They are somewhat distinct. They are increasingly distinct. Will new models arise? Maybe. It is all a question of employment down the line. Employers want to hire young people, but they cannot interview for days thousands of people. It is too expensive. That's the core issue. We need filters.

Cognitive radius: The most interesting change AI brings to software work is not faster coding. It is the expansion of context or range... what I call radius. For decades the job was defined by the limits of what one person could hold in their head. You tended to specialize. You got good at one or two systems, one or two types of problems because keeping more than that coherent was expensive. Context switching hurt. Onboarding to a new project took days or weeks. Simply issuing a pull request to a new open-source repository took days. You had to learn the conventions. Learn the code base. Find out how to test it. Find out how to run benchmarks. With a good model and the right tools, you can load a large amount of relevant information into the session: the structure of several repositories, the conventions of different teams, recent issues, design decisions, performance constraints, the history of related libraries. You can ask questions across them, generate changes that stay consistent with local style, and keep more projects alive at once without the usual collapse in quality. The unit of work stops being “the project I am currently deep in” and becomes something closer to “the set of projects I can keep coherent right now.” This is not the same as having an assistant write code for you. It is more. The gain comes from the ability to hold and operate on a larger working set. You still decide what matters. You still check the results. What changes is how much of the surrounding system you can bring into focus at the same time. The practical consequences are already visible for people who treat this as a skill rather than a novelty: •You can enter a codebase you have not touched in months and become useful again in an hour instead of a day, because the model can reconstruct the local idioms and the reasons behind them from the history and tests. •You can maintain consistency across languages or frameworks. The cost of switching is lower when the model carries the surface details. None of this is automatic. The model does not magically expand your capacity. You have to learn how to feed it the right context, how to structure the questions so the answers stay grounded, and how to verify the parts that matter. The people who get the most out of it treat the model as a high-bandwidth interface to a larger set of information. Most discussion of AI in programming still focuses on the speed of generating individual functions or on the fear that juniors will not learn. Those are real topics, but they miss the larger shift. The constraint that shaped the profession for a long time, the limited amount of context a single developer could keep active, has loosened. What you can usefully own has grown. Learning to operate at that new scale is the part that is still under-discussed and under-practiced. The cognitive radius of a engineer is the distance (in projects, technologies, or time) over which they can still produce consistent, well-informed work. AI has increased that radius.

What is fascinating about what is happening in France, and throughout much of the West, is how reason is gone out the window. It is no longer viewed as a virtue to be someone who thinks. They burn down their schools, demanding more, without thinking about where the 'more' will come. If everyone gets to do a doctorate, paid by the state, and if standards a low enough that graduates neither think well, nor can they serve anyone well... the unavoidable outcome is that we get the level of prosperity that the technology can afford us... not a iota more. We really don't need that much schooling. And we certainly don't want high levels of taxation. Here is part of an essay by professor Emmanuelle Hénin: « Never yet do the unions think responsibly about the means. While France is crushed under an abyssal public debt, they retain the reflex of always demanding more from the State. They mobilize against the slightest increase in registration fees, even for foreigners. Enrollment in a bachelor’s degree costs €178 for the student and €13,000 for the taxpayer—incidentally, the cost of this sabotaged semester amounts to €22 billion. Moreover, if the students who can afford it paid more, they would better appreciate the value of their studies and be less inclined to skip classes or vandalize buildings, as has happened on several campuses. [...] Ideology possesses the admirable faculty of never being contradicted by the facts. [...] Egalitarianism can wreak havoc in France, block social mobility, cause massive failure, gigantic frustration and record youth unemployment; egalitarianism remains the best of policies… [...] That is why, while one may rejoice that the high-school and student movement has exposed fifty years of lies and blindness, one may doubt that the cultural and political conditions are in place to put an end to it. As long as the inspectors of the National Education continue to sanction middle-school teachers who dare to teach the conjugation of the simple past or propose to their pupils texts containing “useless words” (sic), we will not recover. »
Bien difficile de ne pas être d'accord avec le diagnostic d'Emmanuelle Hénin sur la crise de l'Université française. Extraits : "[...] La stratégie est payante, et régulièrement appliquée depuis dix ans : empêcher les cours de se tenir en présentiel, puis arguer du caractère dégradé du distanciel pour obtenir l’annulation des examens ou exiger la moyenne pour tout le monde. Comme le mouvement a commencé deux semaines après la rentrée, les révisions seront légères et les étudiants Erasmus rapporteront de beaux souvenirs de leur semestre à Paris ou à Rennes. [...] Ces revendications expriment un malaise bien réel face à la prolétarisation galopante de l’Université française, au sous-financement de la recherche et à la précarisation du personnel : il est évidemment scandaleux qu’un chercheur en postdoctorat, après huit à dix ans d’études très exigeantes, gagne 1 800 € par mois et que les vacataires soient payés au lance-pierre. C’est pourquoi on réclame des moyens, des moyens et encore des moyens. Mais tous les milliards qui seraient versés dans ce tonneau des Danaïdes ne régleraient pas les problèmes structurels de l’université. [...] S’empressant de se mettre à la remorque de la jeunesse, les instances universitaires - du CNU (Conseil national des universités) à de nombreux présidents d’université et UFR - affichent leur soutien au mouvement et condamnent la « répression » dont il fait l’objet. Tous invitent les étudiants et les professeurs à se joindre aux grèves et à ne pas faire cours, même en ligne, les jours de mobilisation afin de permettre à tous d’aller dans la rue. [...] Jamais, encore, les syndicats ne réfléchissent de façon responsable aux moyens. Alors que la France croule sous une dette publique abyssale, ils conservent le réflexe de demander toujours plus à l’État. Ils se mobilisent contre la moindre hausse des frais d’inscription, fût-ce pour les étrangers. L’inscription en licence coûte 178 € à l’étudiant et 13 000 au contribuable - accessoirement, le prix de ce semestre saboté s’élève à 22 milliards. En outre, si les étudiants qui le peuvent payaient davantage, ils mesureraient mieux la valeur des études et seraient moins enclins à sécher les cours ou à dégrader les bâtiments, comme c’est arrivé sur plusieurs campus. |...] L’idéologie possède l’admirable faculté de n’être jamais contredite par les faits. [...] L’égalitarisme peut faire des ravages en France, bloquer l’ascenseur social, provoquer un échec massif, une frustration gigantesque et un chômage record des jeunes, l’égalitarisme demeure la meilleure des politiques… [...] C’est pourquoi, si l’on peut se réjouir que le mouvement lycéen et étudiant ait fait éclater cinquante ans de mensonge et de cécité, on peut douter que les conditions culturelles et politiques soient réunies pour y mettre fin. Tant que les inspecteurs de l’Éducation nationale continueront à sanctionner les professeurs de collège qui osent enseigner la conjugaison du passé simple ou proposer à leurs élèves des textes contenant des « mots inutiles » (sic), nous ne nous relèverons pas." lefigaro.fr/vox/societe/emma…

How sure are you that the CEO of Nvidia slept with a Chinese spy and gave China access to his tech? I asked grok. I am 100% sure he slept with a Chinese lady. Is she a spy? Maybe. Did he do the communists favours? How do we know? Nvidia sold GPUs to Singapore… some of which may have ended up in Chinese hands. I want to understand. To be clear, my christian values apply: don’t cheat on your wife… it is bad.

There is more C and C++ than you think in Python.
learntToCode @learntToCode Nov 14

More juicy information regarding the latest AMD chips (Zen 6). - Each core can do four 256-bit vector loads per cycle. AMD is ahead technologically with its ability to sustain four loads per cycle and Zen 6 makes this even better. - AMD cut the integer multipliers from three to two, since the third did little. It is uncommon for a processor to be able to do 3 multiplications per cycle. Recent processors from Intel and AMD can... but, apparently, AMD is backtracking. - Recovery from branch mispredictions is faster with no loss in accuracy - VNNI, the AI dot-product instructions, are faster. You can now sustain almost four 512-bit VNNI instructions per cycle. That's a lot of compute. source: chipsandcheese.com/p/talking…
Hello you fine Internet folks, I sat down with AMD’s Mike Clark to talk about Zen 6, the evolution from Zen 5 and what some of the differences are between Zen 6 and Zen 6c and what drove some of those changes. Hope y’all enjoy! chipsandcheese.com/p/talking… youtu.be/AdyPoLBp4No


Whenever I talk to someone in a leadership position in tech… they say much the same things regarding AI and coding… “It is moving so fast.” “We see young people accomplishing unthinkable feats.” “What is going to happen to ordinary Joes? They’ll be gone in two years.” And then you turn around, and you see esteemed academic leaders in computer science who state upfront that they ‘don’t use AI’. In many ways, we are living through sharp bifurcations. That’s expected if you believe in the concept of a technological singularity as a rate of change so fast that people cannot keep up, cannot think through. And that’s what strikes me the most. Not only is it difficult to think about AI and what it means for us, for work… But most people seem to just not think about it at all… acting as if it were obvious, irrelevant, or just non-existent. It seems obvious to me that the few people who decide to keep thinking will just end up surging ahead in due time. There will be a replacement of the elites. AI calls for new elites.

Yes, young people did better in 2025. I have a theory as to why :
U.S. teens saw notable declines in self-reported sadness and suicidal ideation from 2021 to 2025. Relative to 2021, they were also more likely in 2025 to: 1) play sports (49% to 58%) 2) do strength training (45% to 51%) 2) eat fruit >=twice/day (19% to 28%) 3) eat at least one vegetable/day (55% to 63%)

Perfloop found a dangling stack pointer in the Go runtime. It made recover miss a panic. In September, Go tip made recover work inside range-over-func loop bodies. The loop body keeps the address of the panic record to do that. But that record lives on the stack, and Go moves stacks. So if the loop body escaped to the heap and the stack got copied, the address went stale. recover returned nil and the panic kept going. Perfloop filed it with the cause and a small program that fails on tip. The Go team marked it NeedsFix. @cuonglm_ found it also happens on a shrink and sent the fix the next day. David Chase and Michael Pratt reviewed it, and it's now merged.

AMD keeps its microarchitecture names simple: Zen 2, Zen 3, … Zen 6. Zen 6 is new and most of us still lack access. Chips and Cheese measured what changed from Zen 5. Zen 6 sustains 5.96 64-bit adds per cycle versus Zen 5’s 5.25 (same six ALUs). (Yes! Your processor can do that many additions per cycle!) It also recovers faster from branch mispredictions: roughly 3–3.5 ns extra versus 3.5–4 ns on Zen 5. That's an important gain because lots of branchy code is limited by branch misprediction penalties. Overall, Zen 6 processors today can match the best of the previous generation (Zen 5) running at high clocks with more cache. So we might get an realistic performance boost of about 20% with comparable chips. Source: chipsandcheese.com/p/amds-ze…

If someone wants a challenge... the fast_float library is part of the V8 engine (chromium browsers). It calls the library deep down after some processing. In theory, it should be possible to create a variant of the library that is tuned for V8 usage. I have opened an issue: github.com/fastfloat/fast_fl…
How do you make a number parser even faster? Maybe by telling it what kind of numbers to expect? Read my article to find out more...


How do you make a number parser even faster? Maybe by telling it what kind of numbers to expect? Read my article to find out more...
