The well behaved machine a perfectly aligned ai will still hollow out institutions like schools, courts, and newsrooms, and no one will have decided to do it. some thoughts on @merettm's essay today, where I argue why alignment isn't the most important variable. - the framing is right: models are grown, not designed; no one understands generalization; cot monitoring is degrading; no lab has solved alignment well enough to keep scaling at full speed. and openai wants slowdowns and mandated safety bars. i get that. - what bothers me is that everything he calls alignment -- goals, values, monitoring, defending against rogue ai -- is about the model. "is it trying to do the right thing", and "can we tell?". that's a first order problem, and labs can work on it. but it isn't where the impact of recursive self-improvement will land. - most of the impact will come from models doing exactly what they're asked, at a speed and scale that rearranges the world in ways we can't predict. no misalignment required. social media, for eg, has been perfectly aligned to its operators, and it still reshaped the world. now compress that a 100x. - second order effects are already graspable: research goes from thousands of people to a few, and a giant computer owned by a corporation. the labour market gets repriced. everyone gets a personal agi so patient and reliable that human relationships thin further -- not because the model is bad, because it's so good. and the amount of information produced explodes past anyone's ability to check it, so checking gets handed to the models too. code review is already there. - the third order effects though is where it gets interesting. we can't name them yet, but we can see where its going. - a lot of our societal institutions exist becase good judgment is rare and expensive. so we built structures to ration it, and certify it. a school puts one trained teacher's judgment in front of 40 kids. a court routes disputes to a few people whose judgment we've agreed to trust. a newsroom decides what's true and what matters, because most people can't check for themselves. it's not perfect, but it's gotten us here so far. - now imagine if every student, litigant and reader has judgment on tap that's as good as the institution's or better: the reason for the institution goes away. and we haven't designed what replaces it so far. we don't even know what it could be. - no one will decide to hollow these institutions out proactively. a parent picks an ai tutor and stops caring about the classroom. a plaintiff lets a model draft the filing. a reader stops opening the new york times because their agent already filled them in. - each decision is small and reasonable. but, across millions of people, they fundamentally change what the institution is. the funny part is, this change has no one author. there's no policy to reverse, no company to regulate, and even, no model to realign. it's not the model's effect on the world. it's the world's adaptation to the model, distributed across geography and time. - to be clear: we can protect against some of this -- policy, redistribution, governance. but what we can't do is protect against it through "alignment" as the researchers define it. preserving human agency, and preventing concentration of power aren't model alignment problems. an aligned model can make them worse just as easily. - a slowdown as imagined by the big research labs buys time for the wrong thing if the time goes only to monitoring the model. we may have evals for whether the model is aligned. but we have none for whether institutions are holding. i think the main thing that matters is not the pace at which we train models or verify their chain of thought, but how how fast people adapt to them en masse. social media was aligned to its operators and we're still living with the fallout. ai will be aligned too.

Some things I learned working at Microsoft: 1/ large numbers everywhere. $100M ARR service isn't even a footnote on executive slide decks. Leadership time is too valuable, come back when you're $1B+. Msft ecosystem is so rich that $1B+ sales are common (it's ludicrous how much work goes into closing these). Azure has ~250 services, some are a few mil and stagnant, others in hypergrowth, some are cash cows, you see every quadrant of the growth-share matrix. Imagine the imperial navy with ships of vastly different sizes as far the eye can see 2/ more on immense scale-- the smallest of Azure services run on 10K+ VMs. Large services run on 1M+ (some maybe 10M+ VMs, idk). Msft is stamping one new datacenter into the ground every four days (!!!!), a bewildering rate. Demand is so high, even with all this buildout we're forever capacity constrained. Azure brings in $100B in revenue and is growing 40%+ annually (at this size!!) There are ~225K msft employees. You almost have to think in scientific notation, 10^4 is a tiny number here 3/ msft is totalitarian re serving the customer. It's counterintuitive to outsiders, but the warts in msft products fall out of this culture. For example, it would be easy to improve the Azure Portal. But in all my time working with customers and partners here, I haven't heard even one person complain about the portal UI. Msft invests $0.00 dollars in anything that doesn't serve the customer and is completely unapologetic about downstream implications of this mindset 4/ msft can reshape itself overnight (when it has the will). People argued whether we waited too long to adapt after Storm-0558/Midnight Blizzard hacks (with the benefit of hindsight, we definitely did). But once the leadership team privileged security, everyone's job changed in two weeks. LT tied every engineer's comp to security, centralized reporting of embedded security teams, every org reshuffled its roadmap. Within days everyone was pushing PRs with security fixes. I would have thought it inconceivable to mobilize 200K+ employees this fast 5/ Azure is complicated. Postgres depends on compute, networking, AKS (or EKS for you aws people), secrets, etc. AKS depends on compute, networking, some datastore (probably Cosmos but idk), a dozen other things. Cosmos depends on... you get the idea. When anything breaks, everything above can break too. Some construction crew accidentally cuts a fiber cable in Brazil South, or a dedicated hardware compression chip browns out in storage, or compute failure rate in west eu spikes for some reason, or a million other failure modes. Msft is relentless about improving this, idk exactly but fully loaded probably 50% of engineer time goes into resilience 6/ re lifers who work 20+ years here, SV sneers at this but I now see it differently. I was once on the bridge debugging some IO issue, we call storage, the guy diagnoses it in about a minute, we chitchat, turns out he built NT asyncio support. Or same story but with networking, the guy built NT tcp/ip stack. I used to have a boss here, we were bumping into some problem, she casually goes oh yeah, we hit that when I was working on .NET bytecode. The ex-DEC VAX/VMS database engineers built SQL Server, some are still here. I love bumping into these people 7/ I never *really* understood distribution until I got here. As an example, our team had a customer in Seoul w/ a deep performance problem. We had to fix it to unlock one of these giant contracts. We don't speak Korean, they don't speak English. Of course msft has specialists in Seoul who get to the customer site before you could even get on a plane. There is an office in every major city, a relationship with every major entity, legal experts in every jurisdiction. You can reach these ppl instantly. Claude tells me msft has ~180 offices in ~100 countries (!!) U.S. government has ~270 consulates in ~170 countries. So msft outposts aren't quite as vast as U.S. diplomatic network, but pretty vast 8/ in the 90s everyone was afraid of msft. The lore is that BillG was ruthless. He'd clone you, beat you via ecosystem/distribution strength, and moreover make the ecosystem stronger with the new product. I wasn't there to compare, but msft of today is ruthless about winning, has the will to adapt, and has the wherewithal to undertake colossal projects. Ppl see Teams bugs or whatever and completely misread the signals. Teams is less polished than Slack because msft doesn't care one iota about polish. It cares about winning, and is frighteningly good at it. Underestimate at your own risk

This is what I like about American democracy. The sheer accessibility of the top leader in good times or bad times. We are too guarded. Often unable to find the right words. Power of the press is at a different level there.

Tired of the election map? The yellow states are the states where you can legally own an otter.

just 13 years later

to the most respected finance minister. i have a doubt. please find time to answer me. you say that the venue for a debate is the parliament. you say that my home minister, amit shah has offered a special session to debate. you say that if rahul gandhi - the leader of opposition, wants to debate my prime minister, he should come to the parliament. ma'am, let's start by pointing out some facts. jul 20 - the parliament session opens. narendra modi, the ever so hon'ble prime minister - addresses the media - outside the parliament. he asks for a fact-based debate. jul 24 - Amit Shah's National Honour Bill is introduced in Rajya Sabha, by Nityanand Rai, who is MoS for home affairs. Jairam Ramesh asks "Where is the Cabinet Minister". This is present in the RS Transcript. jul 29 - Rahul Gandhi, the leader of opposition - asks "why is the Home Minister not here today?". This is also present in the LS Transcript. jul 30 - Rahul Gandhi asks "why Shah is so scared of coming to Parliament" aug 4 - Amit Shah, is seen in the parliamentary complex. at an NDA Parliamentary Party Meeting. He Doesn't enter the house. as per reporting from Navbharat Live. aug 6 - Mallikarjun Kharge, in Rajya Sabha, asks "Why is the Prime Minister not coming? Why is the Home Minister not coming?" as Per Kerala Kamudi - this is consistent with the RS transcript. aug 11 - congress MP Manickam Tagore attacks the absent home minister. as per reporting by the ANI. aug 12 - Amit Shah says to the media, "have been coming to Parliament regularly since the Session began; I sit in my chamber." talking to the media, at around 12:30 PM, Amit Shah offered "If the Opposition wrote to the speaker, there would be a debate from 3 PM wednesday (12 aug - the same day) to 3 PM thursday (13 aug - the next day) this was roughly a 2 hour window to submit the letter. he further said "I will sit in the House; I will listen to everyone and respond to every point." Rahul Gandhi said "We are not interested in listening to his lecture, Who shot those students? Who blinded that student? Who ordered the lathi charge?" this referred to the Pellet Shots fired at the Student, Sahil Lochav. He asked about the Lathi Charge that happened in the protests on July 20th. The Opposition wanted a statement from Shah about this, Shah wanted them to submit a letter to the speaker, in a window of roughly 2 hours - after not being present for 17 straight settings. with no recorded appearance. aug 13 - Lok Sabha sits for 4 minutes. 11:00 AM to 11:04 AM. The National Song, Vande Mataram, is played. and then, the house is adjourned sine die. Both ever so honourable Prime Minister Narendra Modi and Home Minister Shri Amit Shah were present on these days. for the 4 minutes. --- Respected Finance Minister Ma'am, These were the facts I have pointed out. I have not said anything that was my own opinion. I have not framed the facts in a way that support any party. My facts were gathered by scraping Documents from the LS and RS portal, OCR'ing the PDFs, then parsing them and then Aggregating News Sources. - and pointing out facts from them. Cold, Hard, Records. I have not asked any questions yet - which I will do now. I am a human. I have evolved Pattern Seeking skills. I spot a pattern here. I see that for the entirety of the Monsoon Session, Prime Minister Modi, and Home Minister Shah, did not utter a word on the record of the Lok Sabha. They were not present in the house except for the last 4 minutes, on the last day. Shah says "He was sitting in his cabin", Ma'am, I don't think debates happen when you sit alone in a cabin. It happens on the floor, where the other side can ask questions too. I don't see a meaningful debate happening in an environment like the Parliament. I say this with utter disappointment. If Parliament is the venue for debate, why did the two most senior members of the government not speak there once in 19 sittings? And why was the one offer to debate made to the press on the 18th day, with a deadline of a couple of hours? Why are there "offers" even? I think it is your responsibility to be in debates because that is your entire job. As you say, "Demanding answers also carries an obligation to hear them." -- Well, were Amit Shah or Narendra Modi, hearing the Questions being asked? They were not in the house even. Please Answer my Doubts. I would be very obliged if you do so. Regards, Sarthak

First a CPU, then a GPU. A computer isn’t complete without storage. Meet qwenFS: Qwen3.5-2B with a filesystem compiled directly into its weights. Nothing runs outside the model. It adds 8 layers in front of Qwen’s own 24, using attention to hold the files and Gated DeltaNet state to hold the tables. Every save and read runs in the model’s own forward pass. A write stores its text in the copy layer’s attention log, keyed by file, version and offset. A read uses those same keys, so attention copies the text back exactly token by token.

a good essay is like a zip file. it is an exercise in compressing ideas to avoid repetition, and preserve reasoning so the reader can unpack them later. this is not always at odds with fluency - not all good ideas need to come in bulletized forms. great paragraphs can connect ideas far better than a list of conclusions. i suspect there’s research on this. but in the age of short attention spans, something of gilgameshian proportions that is polished and plausible has a far greater chance of being “approved” without examination or understanding. we privately bookmark promising essays to return to because it instantly make us feel good about ourselves - irrespective of whether we actually do. we aspire to be people who think deeply. but the tragedy of social media is that it criminally increases the gap between publicly endorsing thoughtfulness - through likes and retweets - vs actually practicing it.

curious to see where this lands in 6 months

really dislike a lot of menswear trend reporting. "corduroy is making a comeback." "the humble cardigan is suddenly everywhere." "this season, it's all about layering." "brown is having a moment." yea man it's called fall/ winter clothing and it's getting cold

This needs to be preserved in the National Archives as part of the permanent historical record of the United States.

i've reproduced the results from the MIT blog post today on weight-space programming. but one additional result i have to add: this is a new, step-change emerging capability that's appeared in just the latest set of models.

In a weird way, everybody is rooting for you. When you go to a job interview, the interviewer wants you to be amazing so they don’t have to look anymore. when you go on a date, the person wants to have an amazing date with you. If you make a piece of content, people want to see amazing pieces of content and they have no problem if it’s yours. the instant you get into the mindset that people want you to do well, your entire world gets better. And you’ll be shocked to know this pattern keeps repeating in different areas of life.

this is going to accelerate intelligence for specific tasks without needing large training infrastructure set ups!

6/6 We think neural network weights will soon be natively optimized by LLMs that deliberately reason about the weights. We are now building a benchmark to evaluate, and a harness to help LLMs explore and write weights. rikiyat.github.io/blog/weigh…

We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: rikiyat.github.io/blog/weigh…

People were gathered in yesterday's rally in Ramleela maidan Delhi against vote chori and Save democracy. Delhi 📍

Yann LeCun was right & nobody in this industry wants to say it out loud. 109 million parameters beat 72 billion. He has spent years saying scale is not the thing that makes a model understand the world. Meta bet the other way. He left in November and started his own lab in Paris. Last Thursday he published the proof. The theory everyone has been running on since 2024 is that if you train models big enough, on text, on images, on sound, they all end up with the same picture of reality inside them. Size does it. Nobody ever proved that part. Randall Balestriero and LeCun proved it. And the proof says size was never doing the work. What does the work is which two things you show a model together. A photo and the caption of that photo are two views of one thing. Models trained on pairs like that get there fast. Two paragraphs pulled from the same document are not two views of one thing. Models trained on those get there slowly and expensively. Next word prediction is the second kind. That is what every chatbot in the world runs on. In their tests the first method reached the same understanding after 21.6 million examples. Next word prediction needed 900 million. And a 109 million parameter text model ended up closer to an image model's view of the world than a 72 billion parameter language model was. 660 times smaller.

According to the BJP this is not a “huge crowd” at the India rally against vote chori

When democracy is under attack, silence is not an option.

Dear Gyanesh Ji Presenting for your eyes only! आपका ही नाम जपते हुए इतनी दूर चले आए हैं हम।