If you've ever wondered why smart TVs are so cheap, this is why.
@log4j rooted his Samsung and found an always-on hot mic, automatic content recognition, geolocation, network mapping, ad targeting and user profiling, shared with Samsung's "400 partners".
A lot of it can be blocked in <15 min, wrote a basic guide below.
What your TV is doing
The main culprit is ACR (automatic content recognition). Your TV screenshots what's on screen, fingerprints it, and sends it home to build your ad profile. Samsung does it every 500ms, LG every 10ms [1].
What is insane: it happens on connected devices too.
Plug in your laptop, your console, your work machine, and your TV can fingerprint that via HDMI [1].
He also found that if your Samsung doesn't have a mic, it'll look for one from a connected soundbar or bluetooth device.
x.com/log4j/status/2…
It's on by default, and you likely agreed to it on the setup screen you clicked through in 10 seconds.
Step 1: Turn off ACR
Samsung: Terms & Privacy → Viewing Information Services
LG: Live Plus
Vizio: Viewing Data
Roku: "Use info from TV inputs"
Sony/Google TV: Samba Interactive TV
Menus vary by model, but those are the names to look for. While you're in there, opt out of interest-based ads, reset your advertising ID, decline every optional privacy agreement, turn off voice recognition, and delete apps you never open.
Step 2: Give smart devices their own network
Put the TV, cameras, speakers and vacuum on your router's guest/IoT network. The logic is simple: these devices rarely get patched. If one gets compromised (getting drastically more frequent with AI), it shouldn't be able to see your other devices or home network.
Step 3: Block trackers for the whole house with DNS-level filtering
NextDNS (cloud-based) is easy and has a solid free plan. Pi-hole is a little bit more work if you prefer self hosting. For simplicity we'll focus on the former.
1. Sign up at nextdns.io (free tier is plenty)
2. Privacy tab: add a blocklist (OISD is a good bet) and turn on Native Tracking Protection for Samsung, LG, Roku, Amazon (see attached image)
3. Setup tab: copy your endpoint
4. Paste it into your router's DNS settings as DNS-over-HTTPS/TLS (or use the IPs + "Link IP" if your router only does plain DNS)
5. Bonus: block outbound port 53 except to your router. Some devices hardcode their own DNS to get around you
Then open the NextDNS logs and watch how often your TV phones home. It's eye-opening. If an app or update breaks, find the domain in the logs and allowlist it.
Why they do this
TV prices are down 97% since 2000 [2]. Most of that is better manufacturing and brutal competition. But once margins on the hardware hit zero, the cost moved to you.
In Q3 2024, Vizio lost $6.7M selling TVs and made $115.8M in gross profit from ads and viewing data [3]. They've said on earnings calls that they keep TV prices low to grow the audience they sell to advertisers. Walmart bought them in 2024.
This isn't a conspiracy theory either. Vizio paid $2.2M to the FTC in 2017 for tracking 11 million TVs without consent. Last December, Texas sued Samsung, LG, Sony, Hisense and TCL over ACR. Samsung and LG have since agreed to stop collecting that data without explicit consent [4].
TL;DR
1. ACR off
2. Ads + voice off
3. Smart devices on their own network
4. NextDNS filtering on the router
Sources:
[1] ACR study: arxiv.org/abs/2409.06203
[2] Cost of US consumer goods: aei.org/economics/char…
[3] Vizio Q3 2024: sec.gov/Archives/edgar…
[4] Texas v. Samsung: texasattorneygeneral.gov/news/releases/…
i own a relatively new (2025) samsung tv.
i kept getting weirdly targeted ads.
i tried reversing the firmware but it was aes encrypted and no keys unless tear it apart and do uart/jtag magic.
zero days it is. root shell on new samsung TVs!
discovered PRIVACY HORROR. write up?
i own a relatively new (2025) samsung tv.
i kept getting weirdly targeted ads.
i tried reversing the firmware but it was aes encrypted and no keys unless tear it apart and do uart/jtag magic.
zero days it is. root shell on new samsung TVs!
discovered PRIVACY HORROR. write up?
Never seen a more flagrant and violent breach of privacy in my life, and probably never will. Doubtful that a privacy violation will ever even be able to compare to the scale of this evil and deliberate attack.
Samsung should be eviscerated for this. Criminal charges please
turns out it classifies what audio it hears, before reporting back to samsung.
sensor preferences, dog-barking and baby-crying event counters, and camera streaming duration are built into telemetry paths.
I'm an American, a U.S. Army veteran, and I live in Europe.
And I'll tell you something that might upset a few people back home.
I want my fellow Americans to enjoy the same rights, freedoms, and quality of life that millions of Europeans already enjoy.
I want Americans to have access to affordable healthcare without worrying about going bankrupt because they got sick.
I want American students to pursue higher education without burying themselves in decades of debt.
I want Americans to enjoy stronger digital privacy laws, meaningful consumer protections, paid parental leave, and guaranteed paid vacation.
I want police officers held to rigorous standards of accountability, where de-escalation is emphasized and deadly force is genuinely a last resort.
I want Americans to understand that freedom isn't just about what the government cannot do to you. It's also about having the security and opportunity to actually enjoy your life.
And before someone starts screaming about socialism, understand something.
European countries are capitalist democracies. They have private businesses, entrepreneurs, private property, competitive markets, and individual liberties. They've simply made different decisions about what protections their citizens should enjoy.
Europe isn't perfect. No place is. But pretending America has nothing to learn from other democracies is ridiculous.
I served my country. I love my country. And I don't believe patriotism requires pretending that America is the best at everything.
In fact, I believe the exact opposite.
Real patriotism means wanting your country to improve, even when that means admitting another country does something better.
I don't want Americans to have fewer freedoms.
I want them to have MORE.
More protections. More opportunities. More security. More accountability. More freedom to live their lives without being financially destroyed by illness, education, or circumstances beyond their control.
And having lived on both sides of the Atlantic, I know that much of this is possible.
We should stop asking why Americans can't have these things and start asking why they're expected to settle for less.
Republicans control the House.
Republicans control the Senate.
Republicans control the White House.
Republicans control the Supreme Court.
Republicans control the DOJ.
This is what they did with two years in control:
trump gets:
- Gold all over the Oval Office
- BILLIONS in crypto for himself
- $400 million dollar Qatari bribe jet
- Millions in media and tech settlements
- Gold bars and diamond rings for tariff relief
- BILLIONS in Pentagon contracts for his sons
- $100,000 per month per person for insider info
- $600 million dollar ballroom at taxpayer expense
YOU GET:
- A war in Iran
- Higher gas prices
- Higher utility prices
- Higher grocery prices
- Higher housing prices
- Higher healthcare costs
- 21 dead American soldiers
- loss of trust from allies around the globe
- ICE thugs abusing our friends and neighbors
Now they want TWO MORE YEARS in office,
While they actively try to STEAL your vote.
WE ARE NOT POWERLESS.
Your vote can change this.
Here is George Clooney’s response after Trump accused him of being a "Hollywood elite."
"Here’s the thing: I grew up in Kentucky. I sold insurance door-to-door. I sold ladies’ shoes. I worked at an all-night liquor store. I would buy suits that were too big and too long and cut the bottom of the pants off to make ties so I’d have a tie to go on job interviews. I grew up understanding what it was like to not have health insurance for eight years.
So this idea that I’m somehow the “Hollywood elite” and this guy who takes a shit in a gold toilet is somehow the man of the people is laughable.
People in Hollywood, for the most part, are people from the Midwest who moved to Hollywood to have a career. So this idea of “coastal elites” living in a bubble is ridiculous. Who lives in a bigger bubble?
He lives in a gold tower and has twelve people in his company. He doesn’t run a corporation of hundreds of thousands of people he employs and takes care of. He ran a company of twelve people!
When you direct a film you have seven different unions all wanting different things, you have to find consensus with all of them, and you have to get them moving in the same direction.
He’s never had to do any of that kind of stuff. I just look at it and I laugh when I see him say “Hollywood elite.” Hollywood elite? I don’t have a star on Hollywood Boulevard, Donald Trump has a star on Hollywood Boulevard! Fuck you!"
- George Clooney
actor, philanthropist, humanitarian & activist
Those who are opposed to corporations and favor government do not realize that government is simply the largest, most monopolistic corporation of all and it has a monopoly on violence.
Made a video with @ycombinator on how to launch.
I hope this helps someone out there finally get their first 1000 views, first customers, whatever. This is a game of consistency, not one time viral hits.
TLDR: Record a shitty demo, and just start showing your work.
There’s a misconception that a successful launch video needs to cost tens of thousands of dollars and feel like a high production commercial. But @FarzaTV has a different approach.
Farza is the founder of @heyclicky and previously built Buildspace, one of the largest online schools for people working on their own ideas. He’s been making videos for over a decade, and today his posts regularly get millions of views. Rather than treating launches as rare, high-stakes moments, he’s made them part of the everyday process of building a company.
Farza sat down with YC’s @aaron_epstein to share his playbook for turning demos into stories people actually want to watch. He breaks down how he scripts videos that still feel off the cuff, why founders should launch constantly, and and how to figure out which channels actually convert.
01:44 — Launch Constantly
03:38 — Start Shipping Demos
05:02 — Show the Process
06:07 — The Clicky Launch
07:43 — Write a Hook That Stops the Scroll
10:30 — Why Farza Scripts Every Video
12:04 — Designing the First 10 Seconds
15:15 — Demo-Driven Development
17:25 — You Have to Take a Thousand Shots
20:48 — Farza’s Formula for Videos That Convert
23:01 — Turning Demos Into Stories
24:53 — Show What You're Actually Building
29:11 — Finding Customers on Social Media
31:02 — Which Platforms Actually Convert?
39:01 — Why Farza’s First Hire Was a Filmmaker
40:49 — The Video Stack
45:41 — The 30-Day Launch Challenge
Ben Horowitz's message to early-stage founders on betting on yourself, the doubts that do the most damage, and why fear means you're doing it right:
"Your job is to make your bet pay off. You believed in yourself enough to do it. There are gonna be many times when you question that bet and you question that belief, but that gets you nothing."
"The doubts are gonna be there. Ignore the doubts. The worst doubts aren't the ones people put on you. It's the ones that you put on yourself. Every entrepreneur has them."
"I spent a lot of time with Mark Zuckerberg when he was 22 years old and he had a tremendous doubt about whether he was gonna make it."
"It's not what you feel. It's what you do. You're gonna feel scared. If your guts aren't boiling, you're not even trying at this thing."
@bhorowitz w/ @speedrun GP @andrewchen
Zooming out, GPUs are still renting ahead of a standard 6-year depreciation schedule
Rental price vs. schedule:
- H200: +143%
- B200: +111%
- RTX PRO 6000: +73%
- H100: +54%
- A100: +9%
Most of the gap opened up after April, when token demand took off
Charts of the Week: a16z.news/p/charts-of-th…
GPUs are aging like fine wine, with rental rates up across the board since January 1st
Approximate change YTD:
- H200: +160%
- B200: +120%
- RTX PRO 6000: +65%
- H100: +35%
Even the 6-year-old A100 rents for more than it did at the start of the year
Charts of the Week: a16z.news/p/charts-of-th…
One answer to what's driving GPU rental prices:
Token demand on OpenRouter is up ~27x since last November, from ~6T/week to ~165T/week
Open-weight models now handle ~72% of those tokens, up from ~28% last November
Charts of the Week: a16z.news/p/charts-of-th…
Zooming out, GPUs are still renting ahead of a standard 6-year depreciation schedule
Rental price vs. schedule:
- H200: +143%
- B200: +111%
- RTX PRO 6000: +73%
- H100: +54%
- A100: +9%
Most of the gap opened up after April, when token demand took off
Charts of the Week: a16z.news/p/charts-of-th…
I always thought it was a shame that modest default stack setups made alloca() a generally-frowned-upon technique in C/C++. With 64 bit architectures, give every stack a virtual address space equal to the physical memory by default!
You can do a surprisingly good job of memory management with just the stack, and I contend that it has both correctness and performance advantages.
Allocations are a trivial pointer adjustment, and frees are literally free. Keeping everything linearly packed is friendlier for all layers of the cache hierarchy, and it makes keeping everything in huge / gigantic pages easy.
Leaving temporary intermediate blocks on the stack can seem wasteful, but unless you are dealing with algorithms that make > O(n) temporary allocations, it is usually not a big deal.
Static analysis can check if you store a stack pointer into a structure deeper in the stack, which is a use-after-free, and then memory management correctness is guaranteed.
The discipline of caller-allocates-return is awkward in some dynamic cases. It would take a change in low level calling conventions, but you could plausibly have functions that return with the stack extended from where it was when they were called. I’m not holding my breath for that.
It isn’t obvious if having AI coding agents makes discussions of software architecture more or less important!
"The pace, and the expectations that came with it, felt almost as jarring [in China] compared with America as America's pace does compared with Europe. Everyone was just on it, all the time."
I just edited a bunch of family photos from the Japan and China trips on this. All my Lightroom presets worked beautifully. A few minor things left, but we are 97% there.
So I have zero use for the Adobe Creative Suite sub now. Guess I'll just pay the $600 break-up fee 😂
Slop machines know no rest.
v0.4.5 ships major improvements to lens corrections.
RAWmakase now applies lens corrections embedded in RAW files by cameras: distortion, chromatic aberration, and (for many models) vignetting.
No need to import Adobe *.lcp profiles anymore - but you still can, if you want.
Full release notes: github.com/pch/rawmakase/…
Yesterday's release (v0.4.4):
github.com/pch/rawmakase/…
If you notice any bugs, open a detailed issue on GitHub. Don't complain here 😉
Next, back to work on:
- built-in film simulations and presets
- built-in camera profiles
- missing file management tasks
The highlights/shadows parity research agents have been going in circles, making minimal progress. Need to try a different approach. Or wait for Opus 6.5.
If you don't want to wait to experience the magic of HF open-source robots, you should get a Reachy mini, ships much faster than microduck and as much fun! store.pollen-robotics.com
Dario Amodei was so worried about Y2K that his high school teacher still remembers it decades later.
He’s been a doomer for much longer than I thought!
Long before they were soon-to-be billionaires, Dario and Daniela Amodei were students at San Francisco’s most elite public high school. sfchronicle.com/tech/article/d…
Elon Musk is promoting a video that rewrites the history of Nazism to suit today’s far right. Here’s what happened in 1933, who made the video, and why it matters.
Within weeks of Hitler taking power in January 1933, the organised left was being dismantled.
Communists, Social Democrats and trade unionists were rounded up, and Dachau opened in March, initially to imprison political opponents, above all Communists.
On 2 May the Nazis smashed the independent trade unions and replaced them with the state-controlled German Labour Front. A week later, students burned the works of Marx and Kautsky, with a fire oath declaring itself “against class struggle and materialism”.
Dachau’s first murders came on 12 April 1933, when the SS shot four Jewish prisoners: Rudolf Benario, Ernst Goldmann, Arthur Kahn and Erwin Kahn. Benario and Goldmann were young left-wing activists from Fürth.
The very rich were treated differently. Many industrialists had feared the Nazis were socialists, so in January 1932 Hitler addressed Düsseldorf’s Industry Club, playing down his antisemitism and stressing his anti-Marxism to reassure them. Few backed him openly until he took power.
On 20 February 1933, Hitler met leading industrialists, Gustav Krupp among them, and told them: “Private enterprise cannot be maintained in the age of democracy.” It was conceivable, he said, only where people had “a sound idea of authority and personality”. They contributed more than two million Reichsmarks to the election fund of Hitler’s coalition.
Private ownership was preserved for compliant industrialists, company profits rose sharply, and in the mid-1930s the regime even sold state stakes in banks, steelworks and shipping back into private hands.
In 1934, Hitler also had members of his own movement who took its anti-capitalist rhetoric seriously murdered during the Night of the Long Knives.
Yes, the Nazis controlled prices, wages and investment, as Britain did in wartime. State control tells you how an economy is run, and the question that matters is whom it serves.
The Nazi state crushed independent labour, protected compliant employers, and harnessed the economy to rearmament and conquest.
That is why historians place Nazism on the far right, including conservatives such as Ernst Nolte, who saw anti-Marxism as fascism’s defining feature.
The word in the party’s name was largely a recruiting device, designed to lure workers away from the real socialist and communist parties. A name settles nothing, any more than North Korea’s makes it democratic.
The man Musk is amplifying is Brivael Le Pogam, the Paris-based co-founder of Argil, a Silicon Valley-backed company selling AI video avatars, who boasts of using AI to win hundreds of millions of views.
His self-described “masterclass” spends minutes on price controls and the voting habits of American professors. It never once mentions the Holocaust. Race appears once, as “fuel”.
He is candid about why. With Hitler on the right, he complains, anyone defending “borders” or “the nation” is “one step away from fascism”. That is “why Milei is far right” and “why people call Elon Musk a Nazi”.
A nationalist politics is to be cleansed by relabelling history’s most murderous nationalist.
He even cites Leo Strauss’s “reductio ad Hitlerum”, the fallacy of condemning an idea because Hitler held it, and then commits it from start to finish. And his test of freedom is simply the free market, which would place Pinochet’s Chile, with its torture and disappearances, on the side of freedom.
So the world’s richest man is promoting an attempt to free the far right from its own history, turning the Nazis’ first victims into props for their persecutors.
History records which side Germany’s richest men chose in 1933. It’s worth noticing who is now keen to blur it.
It’s polished, confident and wrong, and it’s built to divide: anyone who objects is cast as the enemy of freedom. Freedom for whom?
Remember who filled Dachau first.
“The U.S. has benefited from the skills that Indians bring without having had to invest much in their education. Many have earned advanced degrees in America, but the bulk of their training happened in India. A more honest way to see it is that India, along with other poorer countries, has been subsidizing the United States.”
If you hang out with happy, optimistic, and ambitious people, it'll rub off too. That's why pruning your social media algorithms is so key. Don't psyop yourself into a depressive hole.
The world is getting more of its electricity from renewables but less from nuclear power.
If the world is to reduce its carbon emissions and tackle climate change, it needs to move away from fossil fuels to low-carbon power.
There are two key sources of low-carbon electricity: renewables — which include solar, wind, hydropower and others — and nuclear power.
While rapid growth in solar and wind has increased the amount of electricity coming from renewables, a lack of enthusiasm for nuclear means it’s playing a shrinking role in the global mix.
In the chart, you can see the share of global electricity coming from fossil fuels, renewables, and nuclear since 1985. Since 2000, nuclear and renewables have followed very different trajectories. Back then, both accounted for a similar share of global electricity. But in 2025, renewables made up more than three times as much: 34% compared to 9%.
The total amount of electricity produced by nuclear plants is about the same as it was two decades ago. But because the world produces much more electricity overall, its share of the electricity mix has declined.
(written by @_HannahRitchie and @parriagadap)
"Yes, men dominate computing. There are many reasons for that; none of them justifies collective guilt. Turning the percentages into a struggle session is nonsense."
I visited Wurster Hall at UC Berkeley today. It was designed by, and currently houses, the architecture faculty.
The UC Berkeley website notes that "Wurster is often voted Berkeley's ugliest building."
I visited because I was curious about whether it would in person seem as unattractive as it does in photos. And, in short: it does. It's gaunt, austere, and filmed in grime. On an elysian campus, it stands out for its bleak hostility. It reminds me of the German submarine base in Bordeaux.
The UC Berkeley page agrees that it was never designed to be attractive, quoting the curator of their design archives: “Wurster and his handpicked architects didn’t want anything romantic, or sentimental — they hated the idea of ‘making pretty buildings. They wanted a serious building.”
I guess we have to concede success relative to their stated goals, but it does make one wonder what it all says about architecture as a discipline.
The Co-Founder of FreeBSD, Jordan Hubbard, has taken FreeBSD… and put Omarchy on top.
He calls it “Daemonarchy”.
And, as expected, the Leftists of Open Source are not pleased. Because it’s a fun looking project. And, presumably, they hate fun.
They’ve already begun calling the FreeBSD founder “fascist”, “fossbro”, and an “a**hole”.
github.com/jordanhubbard/…
억만장자 줄리오 헤레라 벨루티니가 자신의 은행을 조사하던 관리를 해임하도록 푸에르토리코 주지사에게 뇌물을 제안한 혐의로 기소된 뒤 트럼프 사면을 받았고, 그 과정에서 가족이 트럼프의 전 변호사 앨리나 하바에게 거액의 일자리를 제안했다며 사면 절차를 공개해야 한다고 주장한다.
We’ve never seen corruption like this.
Under Trump, a whole pardon industry has grown up around the White House, with wealthy criminals paying well-connected insiders for a shot at clemency. Some have paid up to a million dollars. This case blows past all of them.
A billionaire banker named Julio Herrera Velutini was charged with trying to bribe Puerto Rico’s governor to fire the official investigating his bank. Herrera wanted his record wiped clean, so he went looking for a pardon from Trump.
Herrera’s daughter gave $3.5 million to a group supporting President Trump. Then the Herrera family offered Alina Habba, the President’s former personal lawyer, a job guaranteeing her more than $50 million over four years, at a company with no public profile. Habba signed the deal while she was still a federal prosecutor.
A few weeks after Habba started, Trump gave Herrera a full pardon.
No matter your party, we can all agree that justice should never be for sale. That’s why I helped start the End Corruption Caucus in Congress, and why I’m fighting to bring the pardon process out into the open.
Either we will end the corruption, or the corruption will be the end of us.
China has a Programmers' Day? No wonder they're thriving. The focus on engineering there is unbelievable. So many of the top companies are run by engineers. That's also why they're getting it with both AI and Omarchy so quickly. The West could do with fewer MBAs at the top!
Today is 10/24, Programmers' Day in China (1024 = 2¹⁰) 🎉 So I donated $1,024 to Omacom to support Omarchy. Fun fact: some countries celebrate on the 256th day of the year, and Omarchy's patron tier is literally "256B". Perfect fit, and I'm near the top of the list 😄
In the agent era, maybe it's Makers' Day: anyone who builds real things with code and AI, not just people who write code. Omarchy powers makers 🛠️ @dhh
今天 10.24 中国程序员节，给 Omacom 捐了 1024 美金支持 Omarchy。冷知识：有些国家的程序员节是一年中的第 256 天，而 Omarchy 的赞助档刚好叫「256B」，严丝合缝，我还排在前面 😄
新时代，程序员节也许该叫 Maker 节了：maker 不只是写代码的人，而是用代码和 AI 把想法真正做出来的人。Omarchy powers makers 🛠️
#OmarchyPowersMakers#AllYouNeedIsLove
It was a joy to connect with the @alibaba_cloud team in Hangzhou. Their hospitality was next level, and we're going to work together on making the Omarchy China version perfect. No VPN necessary, all the major Chinese apps preinstalled, dedicated ISO. It's going to be great!
I'm going to Columbia to speak to some students, and apparently you need ID and a QR code to visit the campus. They locked it down after the Gaza protests and still haven't reopened it. It's the only Ivy League university that's closed to visitors.
Elon says:
"we will not be able to boss around AI that are smarter than us"....
while having built his fortune on bossing around thousands of engineers and scientists who are way smarter than himself
🤣🤣🤣
Europe has spent generations mastering science, engineering, and manufacturing. That knowledge and expertise can now power AI of its own.
“Let’s work together to build Europe’s AI economy and put AI to work.” — @JensenHuang at GTC Berlin
Roll Tide and full throttle!
An incredible day in Tuscaloosa with @SecRollins, from signing a new agreement between @NASA and @USDA and meeting with students at @UofAlabama to flying over a packed Bryant-Denny Stadium ahead of Alabama vs. Texas A&M.
The same spirit that takes us into the skies will take us back to the Moon and inspire the next generation to take us even farther.
Trump called it the Big Beautiful Bill.
It cut nearly $1T from healthcare — mostly Medicaid — to pay for massive tax breaks for billionaires and the richest 1%.
How beautiful is that?
The bill leaves millions uninsured and forces many rural hospitals to close.
Working families — priced out of healthcare — will suffer and die — all while Trump and his buddies literally laugh all the way to the bank (with YOUR money).
Ready for a #BlueWaveIn10Days?
This week’s Paper Club is all about inference.
For every FLOP spent training a model, roughly six go to serving it each year — and getting that part right has produced over $220 billion in market cap in the last decade.
But inference is much more than calling model.forward. Every token has to move through a complicated stack of routing, caching, scheduling, kernels, memory, and hardware, all while balancing speed, cost, and quality.
We brought together researchers and builders from @togethercompute, @ollama, @luminal_ai, and @sgl_project & @radixark to take us through that stack from end to end. We’ll see how a token gets from an API request to a GPU and back, why the same open model can perform differently depending on who serves it, and how new approaches to compilers and caching are making inference faster and more efficient.
00:00 - Why inference matters more than training
08:24 - Three kinds of inference innovation
12:14 - Dan Fu (@realDanFu): the lifecycle of an inference request
42:18 - Jeff Morgan (@jmorgan): not all open model tokens are the same
1:04:31 - Joe Fioti (@joefioti): compiling models with search, not heuristics
1:30:30 - Liangsheng Yin (@lsyincs): SGLang and memory for the agent era
Your agent can run 100 subagents. Your laptop can't.
Tell your agent to use the Prized CLI and every subagent gets its own computer, an exact copy of your box, all running at once.
Prized Sandboxes are available today.
a16z's Alex Danco and Moses Sternstein on the bear case that GPUs would depreciate fast, what rental rates actually show, and why the data points to scarcity:
Alex: "There is a whole case that people made on both sides of the argument that these things were assets that would depreciate on a schedule. And in the past 18 months, I think what we have learned is that all our assumptions were wrong."
Moses: "GPU rental rates continue to go up. Even as token pricing comes down, you see more and more consumption, and it's causing relative scarcity amongst the GPUs that we have."
"Even for the older GPUs, the A100s, which have been around already for five years, that were supposed to become obsolete... Whether they had 4-year depreciation schedules, 6-year depreciation schedules, they were supposed to be obsolete almost immediately, and that was the original bear case."
"And plainly, they're not... Everybody is priced above what they should be based on their scheduled obsolescence. And that's true of the older models, the A100 and the H100."
@alex_danco@MosesSternstein
Zooming out, GPUs are still renting ahead of a standard 6-year depreciation schedule
Rental price vs. schedule:
- H200: +143%
- B200: +111%
- RTX PRO 6000: +73%
- H100: +54%
- A100: +9%
Most of the gap opened up after April, when token demand took off
Charts of the Week: a16z.news/p/charts-of-th…
I just made a programming language doing exactly this. As many address-space-reserved huge stacks as the program statically needs, with bump alloc, stable pointers, all compile time safe and yet way faster than what you can do in C++/Rust:
▲ Hugging Face just released Carbon-X, an open 14.8B model that reads raw DNA inside a chat and answers in English.
The key idea is 1 decoder for both DNA and text. A hybrid tokenizer keeps Qwen3's text vocabulary and turns DNA into 6-base tokens. There is no separate DNA encoder: it starts from Qwen3-14B, then gets 302B tokens of DNA + text mid-training, SFT and RL.
Across 5 DNA benchmarks it averages 66.7% with thinking off, vs 67.5% for GPT-6.1 Sol at max effort. It leads every model tested on BioReason Coding (74.9%) and NextBase (39.4%). It is a trade-off though: splice-site questions score only 29.8%, and it is not clinically validated.
Weights are Apache 2.0 and fit on 1 x 80 GB GPU. Hugging Face estimates about $300 to score 2.4M ClinVar variants, vs $32K to $4M on frontier models......
Full analysis: marktechpost.com/2026/10/25/hug…
Model: huggingface.co/HuggingFaceBio…
Blog: huggingface.co/spaces/Hugging…
Repo: github.com/huggingface/ca…
GGUF: huggingface.co/HuggingFaceBio…@huggingface@_lewtun@LoubnaBenAllal1@ClementDelangue
Power is one of the AI buildout's biggest bottlenecks, and US capacity to make electrical equipment is shrinking
US factory capacity since 2019:
- Comms equipment: +110%
- Semis: +77%
- Computers: +16%
- Electrical equipment: -9%
Electrical equipment prices are up 69%, the biggest jump of the four categories
Charts of the Week: a16z.news/p/charts-of-th…
One bottleneck for the AI buildout is skilled labor
US manufacturers are working existing staff longer and paying up for new ones:
- Hours per worker are outpacing headcount growth in 7 out of 10 sectors
- Workers who switch jobs are getting ~10% pay increases, about double the ~5% for those who stick around
Charts of the Week: a16z.news/p/charts-of-th…
Since 2019, imports of AI hardware have grown faster than US production in every category
US production vs. imports:
- Computers: +46% vs. +366%
- Comms equipment: +87% vs. +175%
- Semis: +76% vs. +131%
- Grid & electrical equipment: -7% vs. +69%
Charts of the Week: a16z.news/p/charts-of-th…
a16z's Alex Danco and Moses Sternstein on the AI buildout's missing supply response, and the one category where US production is actually shrinking:
Moses: "You would think with this epic run on prices, which is what we've had, that you would see an increase in output. That's natural what supply and demand would say. You see it to some extent, but not really.
"There's really two levers you can pull. There's utilization, you can get more with what you got, and you can build more capacity."
"The real puzzling one, and perhaps the most important one, is electrical equipment, which is like switchgear, transmission stuff, all the stuff that's really necessary to power on all these GPUs and data centers and the electrification of everything."
"I was so surprised I had to quadruple check this. Production has actually declined. We actually make less of this stuff... Prices have gone up astronomically. There's a huge impulse in demand and we're actually making less of it."
@Alex_Danco@MosesSternstein
Power is one of the AI buildout's biggest bottlenecks, and US capacity to make electrical equipment is shrinking
US factory capacity since 2019:
- Comms equipment: +110%
- Semis: +77%
- Computers: +16%
- Electrical equipment: -9%
Electrical equipment prices are up 69%, the biggest jump of the four categories
Charts of the Week: a16z.news/p/charts-of-th…
Sorry for the wait, the episode is here!
Live action with real robots instead of animation (episode 4): "Anger"
Enjoy!
As always, the software of Microduck and Reachy Mini is open source:
github.com/pollen-robotic…github.com/pollen-robotic…
Since 2019, imports of AI hardware have grown faster than US production in every category
US production vs. imports:
- Computers: +46% vs. +366%
- Comms equipment: +87% vs. +175%
- Semis: +76% vs. +131%
- Grid & electrical equipment: -7% vs. +69%
Charts of the Week: a16z.news/p/charts-of-th…
Power is one of the AI buildout's biggest bottlenecks, and US capacity to make electrical equipment is shrinking
US factory capacity since 2019:
- Comms equipment: +110%
- Semis: +77%
- Computers: +16%
- Electrical equipment: -9%
Electrical equipment prices are up 69%, the biggest jump of the four categories
Charts of the Week: a16z.news/p/charts-of-th…
Simplify your commercial model for delivery, scale, and protection with Global Front End Enterprise.
Read the documentation to discover how it unifies Global External Application LB, Cloud CDN, Cloud Armor, and Service extensions into a single package → goo.gle/4jEdUaG
Thank you to everyone who participated in the Cloud Run Hackathon with Product Hunt, and congratulations to our winner, Inferock! 🎉
Learn more about Inferock → goo.gle/48dWw5s
Working with outside collaborators can add more work for a repo owner, from confirming access to changing permissions.
To keep up with manual housekeeping for one shared org, we decided to use Evergreen: a GitHub App built with native GitHub features.
The workflow 👇
@dhh's new `jade-horse` and `year-of-the-horse` themes look absolutely magnificent on a big TV screen with OmarchyTV. 🐎✨
Honestly can't decide which one to keep as my daily driver right now - which one would you pick?
Europe is building its AI future, and we’re all in.
From AI factories and AI agents to robotics and healthcare, Jensen took the stage at #NVIDIAGTC Berlin to show how AI is taking shape across the continent.
Here's a look back at the highlights:
Global fashion icon @AllSaintsLive is giving its teams the skills, data, and creative freedom to craft their own edge — launching an AI Academy that boosted Gemini activity by 600% while using Google’s Agentic Data Cloud to transform its data foundations ↓
杭州首场 Omarchy Meetup · 11 月 15 日
想改变世界？带上你正在做的东西，来认识同类。
还有一个和 @dhh 一起的小惊喜，敬请期待。
Hangzhou’s first Omarchy meetup · Nov 15
Want to change the world? Bring what you’re building. Meet your people.
And a little surprise with @dhh. Stay tuned.
event.tranquillus.life/en/omarchy-han…
Omarchy on that wild Huawei Mate Fold! I just got it hooked up yesterday as just an external display for a laptop. The 18" screen is glorious. Love to see this! Hopefully we can get a full install possible soon too.
Hypothesis: The architects (who were Berkeley faculty) were not very talented and knew it, and since they knew this would be exposed if they tried to make a beautiful building, they made a deliberately ugly one.
The good news is that the bulk of the work was already done by bsky and mastodon! Honey-potting the most miserable people. Just being on X is perhaps the single most effective pruning technique.
NASA’s Inspiration Tour has landed in in the SEC 🚀🏈
Ahead of Alabama vs. Texas A&M, @NASAAdmin met with students and fans outside Bryant-Denny Stadium, joined @SecRollins to sign a new MOU between NASA and @USDA , and will soon take to the skies for today’s flyover!
Validate your skills in the last course in the Deploy Production Ready Agents learning path—part of the GEAR series.
Earn your advanced skill badge by completing labs that cover ADK, A2A, MCP, and Agent Engine → goo.gle/4jDzP1A
This week’s Paper Club is all about inference.
For every FLOP spent training a model, roughly six go to serving it each year — and getting that part right has produced over $220 billion in market cap in the last decade.
But inference is much more than calling model.forward. Every token has to move through a complicated stack of routing, caching, scheduling, kernels, memory, and hardware, all while balancing speed, cost, and quality.
We brought together researchers and builders from @togethercompute, @ollama, @luminal_ai, and @sgl_project & @radixark to take us through that stack from end to end. We’ll see how a token gets from an API request to a GPU and back, why the same open model can perform differently depending on who serves it, and how new approaches to compilers and caching are making inference faster and more efficient.
00:00 - Why inference matters more than training
08:24 - Three kinds of inference innovation
12:14 - Dan Fu (@realDanFu): the lifecycle of an inference request
42:18 - Jeff Morgan (@jmorgan): not all open model tokens are the same
1:04:31 - Joe Fioti (@joefioti): compiling models with search, not heuristics
1:30:30 - Liangsheng Yin (@lsyincs): SGLang and memory for the agent era
Introducing Pomerado (YC F26): the faster, cheaper, and more reliable alternative to browser use.
Open source. Sign up for free to supercharge your agents 🧵
Today, Enact emerges from stealth with a $30M pre-seed to master dexterity through force and touch.
Vision and language are becoming commodities, but robots need more than a visual cortex to master thousands of fine manipulation tasks.
That is why we’re building an artificial nervous system for robots. It lets them sense what’s happening in real time through their own actuators. No fragile sensors, no extra hardware: built to scale and designed for real-world deployment from day one.
Force and touch lets our system learn efficiently and perform tasks faster, so automation finally pays off where we need it most: where value comes from variance, not volume.
We’re proud to be backed by leading investors, AI researchers, roboticists, and founders, including redalpine, Notion Capital, @firstminutecap & @ylecun LeCun, @arthurmensch (Mistral), @AndrewLonsberry (Path Robotics) and @Thom_Wolf (Hugging Face) alongside many more.
Read what people wrote about us in the comments!
And check out our hiring page:
Signe Lin Kuei Vehusheia, PhD, Yann Glauser, Noah Hüsser, Jonas Eichenberger, Yang Liu, Philip Kneis, Manuel Yves Galliker, Carl Wyder, Aleksander Kamiński, Matt Benice
This new UnHerd magazine essay is about the authoritarian politics of AI doomerism.
It explains how it turned philosophical thought experiments into proposals for mass surveillance of researchers.
The synchronous programming language and system approach banishes asynchrony via shared-tick clocks, includes formal verification. One reason Airbus rang rings around Boeing: Boeing’s Frankenstein mix of asynchronous legacy and sync systems.
Another reason: criminal Boeing CEOs.
Still, FE JS perdures (see quote).
You rightly refer to the synchronous programming lineage — Esterel (Berry, Rigault, Marmorat, Gonthier) and Lustre (Caspi, Halbwachs, Raymond, Pilaud, Plaice) — developed at INRIA, École des Mines de Paris, and Verimag; commercialized as SCADE.
Has anyone tried installing @omarchy linux without disk encryption? Using the latest 4.0.6 iso installer, when asked to format the disk you hit Ctrl-C to do a non-encrypted install... and then a bit later I get this error:
By selecting 'Auto' in the model picker, Gemini runs each job on the best-fit model orchestrating across Gemini and Claude models, and other leading private and open models in the future to deliver optimal quality and lower your cost.
Still, FE JS perdures (see quote).
You rightly refer to the synchronous programming lineage — Esterel (Berry, Rigault, Marmorat, Gonthier) and Lustre (Caspi, Halbwachs, Raymond, Pilaud, Plaice) — developed at INRIA, École des Mines de Paris, and Verimag; commercialized as SCADE.
Most voters trust that the midterm ballots will be counted properly, but fears mount that Trump and his Republicans will try to overturn the results. The best defense may be a major electoral loss for Republicans, which will be harder to ignore.
@Abdalla_Nangere Hi Abdullahi, we hear you and want to help. We’ve sent you a private message so we can look into this for you while keeping your account details and personal data secure. Please check your inbox so we can resolve this together. - BP
Verdict: Unencrypted installs aren't broken in general: CI's unattended unencrypted whole-disk install passed today with the same disk config, and no matching reports are queued. The photo's mount failure most likely comes from leftover signatures on that disk, which the whole-disk path never clears with wipefs; blkid -p /dev/sda2 on that machine would confirm it.
Still, we should catch that case!
To be continued. I'll follow up after talking with folks about both issues (goggles not applying to Ask; biased answers that reflect disreputable yet official sources).
After poking at a hosted version of a model in the family we're using, the issue seems to be search grounding and not pretraining or alignment.
The (older) hosted model in the family I tested did show alignment against asking for any race or religion doctor, but without bias.
Still, FE JS perdures (see quote).
You rightly refer to the synchronous programming lineage — Esterel (Berry, Rigault, Marmorat, Gonthier) and Lustre (Caspi, Halbwachs, Raymond, Pilaud, Plaice) — developed at INRIA, École des Mines de Paris, and Verimag; commercialized as SCADE.
Just an FYI, 30 km/hr is 18.6 mph. It's an insanely low speed limit for a neighborhood. Last year in England, I received a camera ticket for going a very slow 27 mph in a suburb that was supposedly posted at 20 (I didn't pay). And to champion this with these guys is dystopian.
log4j@log4j
