Software dev and psychologist, keen on functional programming. Former neuroscientist. Game AI researcher at night. Petty dabbler of the cognitive arts. He/they.
re "openness in AI" discussion: there is a lot of talk about the importance of open-weight models. I want to highlight another topic: openness of APIs.
LLM-access APIs were traditionally mapping in a relatively straightforward way to what is actually sent to the LLM: a sequence of messages, each with a different role, and each call specified the exact sequence. this gives both some level of transparency into the underlying model, and also a level of control of our interaction with the LLM.
both Anthropic and OpenAI are attempting to transition into alternative APIs, ones where the session state is managed on the server. currently they support both versions, but i predict a not-so-distance future where they will be deprecating the old APIs and leaving us only with the new ones.
this will add a layer of obfuscation between us (the API users) and "the AI" (no longer just an LLM, we will now communicate with an opaque system). This will be very bad for science, engineering and progress around AI: we (the public, the research community, the builders) will have much less visibility into what is actually being run, and also much less control on where we can intervene. this means research on agentic-harnesses and context management, for example, will be practically impossible.
this is a bad outcome. we should be able to resist it, or at least to loudly signal against it. i think it is as important as supporting open-weight models, if not more.
how do we organize around this cause?
¿Interesado en PSG infantil?
Este curso, dirigido por la Dra Peraita Adrados, pionera y referente de la M. del Sueño en España, permite profundizar en la técnica e interpretación del PSG infantil, del neonato al adolescente.
¡¡Yo no me lo pierdo!!
https://t.co/iSR1u6TiEQ
I still really do not understand guys how are you living with "I do not write/review/look/read code anymore".
Yes, it is silly to argue that LLMs have not changed software engineering. But really to such degree that you are only prompting? Like, how does it work in reality? I really want to know.
Based on my experience, I have no idea how to leave it without review. It always skips something, mixing things, assumes, forgets....
Like... hey, please build feature X, implement it in steps A, B, C, with these constraints.... And it starts....
- oh yes, you said A, B, C, but I started with B
- I should have checked instead of assuming
- Yes, you are right! You did not ask about D
- and so on and so forth
Did I write code better manually or would write better now without llm? No no, not really. But before it was at least reviewed.
So... how?
Hoy he participado en la XIII reunión de la Asociación de Matronas de Madrid. He descubierto el papel que juega la matrona, más allá del embarazo y lactancia, a lo largo de toda la vida de la mujer.
En mi ponencia he hablado sobre SPI gestacional ➡️🧵
@NeurologiaHRI@HospiLaPaz
Se podrá criticar a la IA por muchas cosas, pero por el nivel de memes sobre el tema Jensen Huang desde luego que no.
Vamos con selección con los mejores del día, dentro hilo:
I wanted to prove FSD is safer than human drivers. Genuinely. I went to the data expecting to find the number everyone quotes: one accident per six million miles on FSD versus one per 670,000 for humans. Nine times safer. Case closed.
The claim cannot be made. Not "the data is weak." The comparison itself does not exist.
Start with the equipment. The "human driver" baseline is every car on American roads, and the average American car is over twelve years old. That baseline includes bald tires, worn brake pads, cracked mirrors, and drivers with no emergency braking, no blind spot warning, no lane keeping. A modern Tesla is a rolling sensor suite with the best crash-avoidance hardware ever sold to consumers. So I went looking for the honest comparison: FSD miles versus miles in comparable modern EVs with the same braking, comparable sensors, comparable warning systems. That dataset does not exist in any usable form. Nobody publishes it. Tesla certainly doesn't.
Then the miles themselves. FSD engagement is self-selected. Drivers turn it on in good weather, in daylight, on roads they trust it with. The owner's manual itself warns the system "may not operate as intended" in rain, snow, fog, and low light, which happen to be the conditions where humans crash most. So the FSD number is computed over the easiest miles in the newest cars with the richest, most insured drivers, while the human number absorbs every drunk at 3am and every teenager in a beater.
And it's supervised. Every mile in the FSD column had a human babysitter whose interventions, whose saves, get credited to the machine. The comparison isn't machine versus human. It's machine-plus-human versus human alone, with all credit assigned to the machine.
Darrell Huff described this exact move in How to Lie with Statistics in 1954: if you can't prove what you want to prove, demonstrate something else and pretend they are the same thing. Rolf Dobelli in "Art of Thinking Clearly" calls the other half of it the swimmer's body illusion: swimmers have great bodies because bodies select swimmers. Teslas crash less because new luxury cars with safety suites driven by affluent middle-aged owners crash less. Selection, not software. When Forbes decomposed Tesla's own 2020 report by road type, the Autopilot advantage shrank to about four percent, inside the margin of error. Tesla has refused every request since to release the breakdown.
So here is where I landed. I wanted the claim to be true and provable. It is neither provable nor disprovable with the data that exists. The honest sentence is "we don't know." Anyone who tells you FSD is statistically safer than humans is guessing. Anyone who tells you it's statistically more dangerous is also guessing. The difference is that only one of those guesses is attached to a stock price, a monthly subscription, and a man telling you the feature will be finished next year, every year, since 2016.
When someone quotes you the nine-times-safer figure, ask one question: compared to whom, driving what, in what weather, measured by whom? The silence is the answer.
Some of my friends think AI will make it possible to build AAA games in just a couple of months. I think that seriously underestimates how complex game development is.
I often point them to game designer Liz England’s “The Door Problem.” It’s a great example of how much thought goes into game design. Making a good game still takes hard work and experience.
#gamedev #game
The dogma of the last decade in computer vision is that all classical algorithms should be replaced by neural networks. This has been very productive but, counterintuitively, with the latest frontier models, I find myself sometimes turning back to classical methods.
If you have the training data, the dogma still holds. But there are situations where data is hard to get, there exist good classical algorithms for solving a problem, optimization methods are just better, or where some simple code just does the thing you want.
When writing code was time consuming and error prone, training everything made sense. Now, that a frontier model can code up and test solutions quickly, getting a solution through coding can be faster than gathering training data and training a model. Consequently, I think there will be more hybrid approaches that blend trained models and code.
We now have generic, high-level, planners that can execute sequences of lower-level tasks to solve large problems. These lower-level tasks may be solved by a neural network, a classical method, or specialized code written on the fly.
It never pays to be religious about technology. The future is probably hybrid. Pick the thing that works for you.
Put a heart ❤️ next to a coding problem, and Claude & GPT write recursive code 100% of the time (0% baseline). They also can't tell you why!
Excited to share “A Heart Means Recursion” from my summer at @GoogleDeepMind, on how VLMs read & communicate with visual symbols. 🧵
When you torrent, you join a swarm. Every peer in that swarm can see every other peer’s IP address, that’s how the protocol works. It has to know where to send the pieces.
Anti-piracy firms join swarms for popular titles and log every IP that connects, with a timestamp. It’s automated, which is why it took just two days.
The firm sends starlink the IP and timestamp. Starlink checks which customer had that address at that moment and forwards the notice. Resellers like sasktel say they don’t hand your identity to the rights holder unless a court orders it. starlink’s policy says repeat violations can lead to “service-related consequences.”
Piracy or not your ISP can see:
Every domain you connect to (DNS lookups, and the site name in the HTTPS handshake)
every IP you talk to
when you’re online and how much data you move
what kind of traffic it is. Bittorrent has a recognisable pattern even when it’s encrypted
HTTPS hides the page and what’s on it.
Yes use a VPN but it doesn’t make you invisible it only changes who can see you. Your ISP sees one encrypted tunnel to a VPN server. The swarm sees the VPN’s IP. The VPN provider now sees everything your ISP used to see. you’ve swapped who you trust.
Your ISP has always had the logs. Paramount just showed you they get used.
> be Sam Zeloof
> decide that making computer chips shouldn't require a billion-dollar factory
> turn your parents' garage into a semiconductor laboratory
> start making individual transistors while still in high school
> then attempt something almost nobody has done at home
> manufacture an actual integrated circuit from scratch
> in 2018, build the Z1 with six transistors
> not buy a chip, not assemble a kit
> actually manufacture the silicon chip yourself
> use chemicals, furnaces, microscopes, and ultraviolet light
> build your own machine to print microscopic patterns onto silicon
> because commercial photolithography machines can cost millions
> finish the entire first chip using 66 steps
> then go back and make something 200 times more complex
> build the Z2 with 1,200 transistors
> use salvaged equipment and machines he repairs himself
> even buy a broken electron microscope for $1,000
> a machine that originally cost $250,000
> design chips, fabricate them, and test whether the transistors actually work
> all from a garage in New Jersey
> for perspective, Intel's first commercial microprocessor had 2,300 transistors
> Sam is using technology inspired by the semiconductor industry of the 1970s
> his chips aren't competing with modern processors
> he's proving that making silicon isn't exclusively a billion-dollar industry's privilege
> imagine what happens when more people can build their own chips
Why isn’t Spanish food more famous?! I’m on my fifth trip to Spain and the dining is amazing. I’ve had one (1) poor meal in those trips (a reheated pizza in a hotel lobby, so not really Spain’s fault).
The cuisine is as good as or better than the famed French & Italian. AND the prices are way lower.
I think it’s because
1) French & Italian have better marketing;
2) many foreign visitors - British, for instance - frequent “touristy” spots on the coast and get fed dumbed down versions;
3) it’s highly regional, so there’s no coherent national style.
4) it’s dependent on super-fresh local ingredients, which you can’t really translate to another country.
5) while the entrees and sides blow me away, frankly the Spanish desserts are pedestrian by comparison. That dings it a bit. Yeah they’re passable but not on the level of the rest.
We had “arroz caldoso de bogabante” which I’m told the restaurant mispelled on purpose. The rice is soupy ( different from paella) and there’s a whole lobster in there. We also had grilled alcachofa for starters.
What are your ideas or theories or experiences about Spanish food?
The year is 1854. Cholera is killing thousands in London. John Snow traces an outbreak to a contaminated water pump. Everyone blames bad air. Eventually cities separate sewage from drinking water. Clean water and sanitation will save more lives than almost anything medicine invents.
The year is 1867. Surgery is often more dangerous than the disease it treats. Joseph Lister starts cleaning instruments and wounds with antiseptic. Other surgeons ridicule him. Infections plummet.
The year is 1890. Your daughter has diphtheria. A thick membrane is forming in her throat, slowly suffocating her. The doctor injects a new antitoxin. She survives. A disease that killed tens of thousands of children will become almost unheard of in America.
The year is 1922. Your 14-year-old son has diabetes. He is starving to death no matter how much he eats. The only treatment is a starvation diet. The doctor injects insulin. His blood sugar falls. Children who would have died within months will grow up and have children.
The year is 1943. Your daughter has pneumonia. Before antibiotics, doctors could do little but hope. Now they give her penicillin. The fever breaks. Infections that routinely killed healthy adults become treatable.
The year is 1955. Parents are terrified of polio. Swimming pools close. Children become paralyzed. Some spend their lives in iron lungs. Jonas Salk develops a vaccine, tested in more than a million children. Within decades, polio disappears from America.
The year is 1960. Your wife is giving birth, and something goes wrong. A century earlier, she might have died. Now there are blood transfusions, antibiotics, safer cesarean sections, and anesthesia. Maternal mortality has fallen dramatically. Millions of women survive childbirth who otherwise wouldn’t have.
The year is 1978. Your blood pressure is 180 over 110. You feel fine. The doctor prescribes a pill. Clinical trials show that treating hypertension prevents strokes, heart failure, and death. You take it every morning. Nothing happens. That’s the point.
The year is 1980. The WHO declares smallpox eradicated. A disease that killed hundreds of millions in the twentieth century has disappeared from nature. Not treated. Not managed. Eliminated.
The year is 1996. Your brother has HIV. A decade earlier, the diagnosis was effectively a death sentence. Now doctors combine antiretroviral drugs. People who had months to live begin living for decades. Many will have near-normal life expectancies.
The year is 2014. Your father has hepatitis C. For decades, the infection has meant liver failure and cancer. Now he takes pills for twelve weeks. The virus disappears. A lifelong infection becomes curable in more than 95% of patients.
The year is 2019. Your daughter has cystic fibrosis. Her lungs are slowly failing. For decades, treatment meant managing infections and delaying the inevitable. Now doctors have a drug that corrects the underlying protein defect. Her breathing improves. Hospitalizations plummet. Predicted median survival rises from roughly 40 to more than 60 years.
The year is 2026. Your doctor has ten minutes and a screen. Your blood pressure is high, so the tablet is for life. Your cholesterol is high, so another tablet is for life. Your weight is high, so she offers an injection. The drugs have side effects. Not everyone needs them. But clinical trials show that appropriate treatment prevents heart attacks, strokes, kidney disease, and death.
None of these treatments is perfect. None needs to be.
For most of human history, parents buried their children, infections killed healthy adults, and childbirth killed young women.
Today, global life expectancy has more than doubled since 1900. Child mortality has collapsed.
Medicine has made terrible mistakes. We should remember them.
We should also remember the hundreds of millions of people alive because somebody got it right.