Brandon Wang

44 followers · 17 following · 2145 views

on the atlas — 116

highlights — 730

During the 30-day pre-release government review period, employees would be limited from accessing models. The models would be stored in high-security environments and there would need to be detailed logs on who is accessing the models. The review process will include various administration officials rather than a single office or agency.
Trump AI framework excludes open AI models
Open models are excluded, and the framework explicitly says nothing in it should be interpreted as restricting open models once they've been released.
Trump AI framework excludes open AI models
On July 7, a 737 freighter, operated by a tiny Pakistani cargo airline called K2 Airways, took off in the late afternoon from Sharjah in the United Arab Emirates and flew east toward Karachi with a five-person crew. Its route took it just south of the Strait of Hormuz, an area that had been experiencing intense GPS jamming due to the US-Iran conflict. Later, after nightfall, the flight crew called Karachi air traffic control and reported a “navigational system issue,” according to the Pakistan Civil Aviation Authority. In the three minutes that followed, the plane dove 5,000 feet, climbed 6,00…
A Civilian Plane Crashed in New Mexico. Was the Military’s Tech to Blame? | WIRED
For a few years, US aviation was spared the disruptions of anti-drone electronic warfare. Then it started to happen here, too. In March 2025, airliners flying into Ronald Reagan National Airport in Washington, DC, received spurious alarms from a collision-avoidance system, and several had to abort their landings. It later turned out that the Secret Service was testing electronic warfare equipment at the vice president’s residence.
A Civilian Plane Crashed in New Mexico. Was the Military’s Tech to Blame? | WIRED
EVEN ABSENT GPS jamming, medevac is one of the most dangerous categories of civil aviation. (Kreindler, a law firm specializing in air crash litigation, says that medevac flights have an accident rate more similar to combat flying than to civil aviation.) Flights are often organized on short notice, and they fly into airstrips that might be unfamiliar to the flight crew, and because human lives are at stake, there is an incentive to fly when weather conditions are marginal.
A Civilian Plane Crashed in New Mexico. Was the Military’s Tech to Blame? | WIRED
At the White Sands Missile Range that night, US military personnel were conducting a GPS jamming exercise that left the King Air pilots—and anyone else within hundreds of miles—unable to use modern navigation systems. Forced to revert to older technology, ones that they rarely if ever use, the medevac pilots got disoriented and crashed into the side of a mountain. There were no survivors.
A Civilian Plane Crashed in New Mexico. Was the Military’s Tech to Blame? | WIRED
What was true in Thucydides day is true in our own. The simplest explanation for modern academics’ hostility to 21st century capitalism’s “structures of power” is their complete exclusion from them.
History is Written by the Losers – The Scholar's Stage
Even winning historians need time in defeat to write their histories—had Churchill’s party not been kicked out of power by British voters after the Second World War was over, Churchill’s famous account of that war would never have been written.
History is Written by the Losers – The Scholar's Stage
Thucydides’ obsession with Brasidas is easy to understand once his personal relation to Thucydides is made clear. His portrayal of Brasidas as daring, brilliant, charismatic, and clever beyond measure also begins to make sense—the greater Brasidas’ past feats appear, the less damning Thucydides defeat at his hands becomes.
History is Written by the Losers – The Scholar's Stage
In his roundtable post, “Treason Makes the Historian,” Lynn Rees lists a few of the type. Herodotus wrote his history only after his exile from Halicarnassus; Xenophon wrote his memoirs only after his faction was forced out of Athens. Polybius was once a general for the Archean League, but wrote his history as a hostage at Rome. The destruction of Judea was chronicled by a Josephus, a Jew.
History is Written by the Losers – The Scholar's Stage
These stages are not a fixed sequence that automatically ends with releasing the model’s weights. Instead, we think of them as a way to gather evidence and build resilience while moving towards greater openness. Progression should depend on what we learn about the model and whether the surrounding ecosystem is ready.
A Safe Path to Open Weights - Thinking Machines Lab
Expanding access in stages lets society build up its defenses as it goes: a release can begin with inference API access for a limited population, widen to monitored general availability, and eventually reach open weights.
A Safe Path to Open Weights - Thinking Machines Lab
For example, we are particularly interested in whether dangerous capabilities that depend on specialized knowledge can be reduced through pretraining data curation or post-training interventions.
A Safe Path to Open Weights - Thinking Machines Lab
While I see a similar amount of low-hanging fruit today as I did a year ago, the efforts (or physical resources, GPUs) it can take to unlock them have increased. This pushes people to keep going one step closer to their limits. This is piling on to more burnout. This is also why the WSJ reported that top researchers “said repeatedly that they work long hours by choice.” The best feel like they need to do this work or they’ll fall behind. It’s running one more experiment, running one more vibe test, reviewing one more colleague’s PR, reading one more paper, chasing down one more data contract. …
Burning out - by Nathan Lambert - Interconnects
I’m shifting to see more limitations on the human capital than the financial capital thrown at today’s AI companies. As the technical standard of relevance increases (i.e. how good the models people want to use are, or the best open model of a given size category), it simply takes more focused work to get a model there. This work is hard to cheat in time.
Burning out - by Nathan Lambert - Interconnects
There’s a famous section of the book Apple in China, where the author Patrick McGee describes the programs Apple put in place to save the marriages of engineers traveling so much to China and working incredible hours. In an interview on ChinaTalk, McGee added “Never mind the divorces, you need to look at the deaths.”
Burning out - by Nathan Lambert - Interconnects
There are many different sizes and types of models that matter, but as the market is now more fleshed out with resources, all of them are facing a constantly rising bar in quality of technical output. People are racing to stay above the rising tide — often damning any hope of life balance.
Burning out - by Nathan Lambert - Interconnects
“OSTP at times is in the room, but Michael [Kratsios] is not seen as an effective executor of policy,” said the person close to the White House. “So they bring him in because OSTP has a lot of staff, and there’s no reason to undermine him. But he’s not always in the room.” Similarly, the person added: “Commerce and Howard are sometimes in the room, but not always in the room. Sometimes these silos start working on things and maybe even produce things without sign-off from the West Wing and without total collaboration across government.”
Inside the White House debate over what to do about Chinese AI - POLITICO
Axios reporting that the administration was considering banning Chinese open-weight AI models caused widespread panic in Silicon Valley, but no such sweeping proposal was seriously on the table at that point, according to two people familiar with the discussions, granted anonymity to discuss private conversations.
Inside the White House debate over what to do about Chinese AI - POLITICO
In the six days since Moonshot AI released its Kimi K3 model, Bessent and Office of Science and Technology Policy Director Michael Kratsios offered diverging policy ideas, none of which so far have been publicly announced by the White House.
Inside the White House debate over what to do about Chinese AI - POLITICO
Labeling data in China doesn’t have a cost advantage—especially on labeling high-end data there’s no cost advantage—which makes it very hard for us to invest in labeling data like the US. This path in China is very hard, because labeling data is simply too expensive—whether we outsource labeling or label ourselves, both are very tough. So right now it’s basically walking on two legs. It’s not that we completely can’t label—rather it’s that labeling data has some low-cost, some high-cost. We label the low-cost ones first.
DeepSeek's Liang Wenfeng Breaks His Silence - by Fred Gao
So we call this “drawing lottery tickets.” The threshold is very low—anyone can go draw, but who can draw out something—this maybe I also don’t know whether it depends on talent or what.
DeepSeek's Liang Wenfeng Breaks His Silence - by Fred Gao
Right, I think embodiment definitely still has to be entered—ultimately embodiment. So for our company, naturally, the endpoint might all be embodiment. Because for a normal person, their needs aren’t computers, right? Because normal people—their eating, drinking, entertainment, clothing, food, housing, transportation—they don’t need computers.
DeepSeek's Liang Wenfeng Breaks His Silence - by Fred Gao
We currently feel that maybe the problem every company faces is insufficient talent. But I think this talent shortage will be phased. In the early period of every industry’s development, talent is insufficient. Including previously making websites—when websites were first being made, there were very few people making websites, talent was very scarce. Later the internet needed server-side work, and talent was also very scarce. But this kind of talent shortage is very quickly resolved—just two or three years, because large amounts of people are cultivated.
DeepSeek's Liang Wenfeng Breaks His Silence - by Fred Gao
Our current resources, and our resources within this year and the coming months, including soon-to-arrive large resources, are only enough for us to do more experiments at the 10B-active scale. Because at the several-tens-of-B active scale, there are still many experiments to do, still many things to figure out. We should be relatively far from being able to train an 800B model—still plenty of time, don’t have that many cards.
DeepSeek's Liang Wenfeng Breaks His Silence - by Fred Gao
Actually, spending this much money is very difficult—you can’t buy that many cards, they’re very hard to buy, and prices are very high, and you can’t spend extremely high prices to buy them—you still have to ensure the price is reasonable.
DeepSeek's Liang Wenfeng Breaks His Silence - by Fred Gao
Our total compute was relatively little last year—this year we’re very aggressively expanding this compute. We currently have roughly 20,000 H-equivalent, and in the coming months we’ll have large batches of machines bought over, basically all NVIDIA.
DeepSeek's Liang Wenfeng Breaks His Silence - by Fred Gao
When dealing with the outside, our attitude is: we only do the AGI main line. This is what I just mentioned—GPT, CoT, Agent and so on—only the main line. The AI field is very broad, there are many things we feel aren’t on this main line—for example, 3D, video generation—I think they may not have much relation to the intelligence main line, so we won’t do them.
DeepSeek's Liang Wenfeng Breaks His Silence - by Fred Gao
At that time everyone was fighting bloody battles over the C-end, and the result was led away by someone who didn’t compete. This also indeed shows the logic I mentioned earlier has merit—there indeed are talent and organizational advantages.
DeepSeek's Liang Wenfeng Breaks His Silence - by Fred Gao
This roadmap is because for each step, the new things you need to do are very few. This roadmap lets us avoid working overtime. But if the roadmap were reversed—say, it needs to first achieve embodied intelligence—then this is very tiring to do ourselves, a very bitter task. We don’t want this kind of roadmap—we want to do it a bit more relaxed
DeepSeek's Liang Wenfeng Breaks His Silence - by Fred Gao
But we chose a very restrained approach—that is, I won’t compete with you over this thing, because there are still watermelons behind, and what’s in front is maybe just sesame seeds.
DeepSeek's Liang Wenfeng Breaks His Silence - by Fred Gao
For example, last Spring Festival we suddenly had many users, but we didn’t pursue retaining these users, or monetizing these users, or grabbing these commercial benefits, cashing in on the users.
DeepSeek's Liang Wenfeng Breaks His Silence - by Fred Gao
Because this is the purpose for which we spent so much effort, so much care making this model well. The purpose is to be very cheap, very effective, letting everyone fully use it. We just feel this is happy, this is our motivation, this is our vision, this is the consensus that allows our company to come together to do this thing—this is our company’s internal consensus.
DeepSeek's Liang Wenfeng Breaks His Silence - by Fred Gao
If there’s a narrative I like, it’s a group of ordinary people accomplishing extraordinary things, rather than a group of geniuses accomplishing extraordinary things. This is closely related to our restraint—it’s of one piece with our restraint and our vision.
DeepSeek's Liang Wenfeng Breaks His Silence - by Fred Gao
This is different from previously open-sourcing a piece of software, because that software’s market wasn’t that big. But AI is simply too big. If we want to monopolize this benefit, then we’re bound to be abandoned by history. I think the main point is that this is an objective law, this is a view of history. It’s not that if I don’t open-source, I can monopolize this market—that theoretically doesn’t conform to objective reality. You’ll definitely encounter many obstacles, there will definitely be other methods to stop you from achieving this goal.
DeepSeek's Liang Wenfeng Breaks His Silence - by Fred Gao
Like much of corporate America, the terrorist groups appear to have teams dedicated solely to working on A.I.
How Terrorist Groups Are Using A.I. to Gain an Edge in Battle - The New York Times
Common topics included managing an account on an A.I. platform, suggestions on generating useful answers and tips on evading safety restrictions.
How Terrorist Groups Are Using A.I. to Gain an Edge in Battle - The New York Times
American intelligence analysts say terrorist groups are also beginning to use A.I. to help 3-D-print weapons parts used in plots, according to a former top U.S. official briefed on the matter. For example, A.I. is helping some of those insurgents with design and manufacturing guidance for drone components, repair parts and munitions fittings, said the former official, who spoke on the condition of anonymity to discuss internal assessments.
How Terrorist Groups Are Using A.I. to Gain an Edge in Battle - The New York Times
“It is like a human robot! We used it a lot.”
How Terrorist Groups Are Using A.I. to Gain an Edge in Battle - The New York Times
“We saw in a movie how motorcycles can jump over bridges,” a former Boko Haram commander told Antonia Juelich, a terrorism and technology researcher at Cambridge University. “We used A.I. to learn how to do this. We gave it information, like what motorcycles we use and the distance we need to jump and so on, and it gave us steps on what we have to do.”
How Terrorist Groups Are Using A.I. to Gain an Edge in Battle - The New York Times
“When I stayed away from the internet, I might have enjoyed things I treasured, but there were also costs,” he reflects. “A slow train means beautiful scenery, but also low efficiency, and the trip might not bring economic benefits. But I’m chasing a life that I personally enjoy most.”
134 Days, 68 Places, Zero Internet: One Man’s Journey Through Digital China
At some post offices, mailing had become so rare that staff seemed unsure how to handle it. Though Yang sent all the letters he wrote, some were lost along the way.
134 Days, 68 Places, Zero Internet: One Man’s Journey Through Digital China
A station worker accused him of being a spy because “only spies don’t use phones, since they are afraid of being tracked.”
134 Days, 68 Places, Zero Internet: One Man’s Journey Through Digital China
“It wasn’t a large railway station, and they weren’t used to doing that manually, since most people change tickets on their phones now,” Yang recalls. “Though they eventually helped me change it, it took a long time. They complained and felt like I was just making trouble for them.”
134 Days, 68 Places, Zero Internet: One Man’s Journey Through Digital China
At his first stop, a chain hotel in Linfen, another city in Shanxi, he couldn’t get a room in person because the hotel only accepted online bookings. The front desk staffer told Yang there was nothing he could do to help him check in, though kindly drew him a map to another location of the same chain that could accommodate walk-in guests.
134 Days, 68 Places, Zero Internet: One Man’s Journey Through Digital China
He had doubts about whether he would be forced to abandon the effort partway through and start using a phone again if circumstances called for it, but he believed the challenges were also part of the experience. They helped him better understand just how deeply digitalization runs through modern Chinese society.
134 Days, 68 Places, Zero Internet: One Man’s Journey Through Digital China
Into Yang’s backpack went several books, two cameras, his ID card, a few changes of clothes, two paper maps, some cash and bank cards, a notebook and a pen, and — in a deliberate nod to the analog world he was about to reenter — several Chinese writing brushes, ink, and xuan paper for writing letters. No phone. No data.
134 Days, 68 Places, Zero Internet: One Man’s Journey Through Digital China
A later hypothesis (Michaud et al. 2023, Brill 2024) assumes that knowledge or skills are learned in discrete chunks (“quantized”) and that the frequency distribution of these skills follows a power law. The model learns common skills first and rare skills later, resulting in a smooth power-law decay in loss.
Scaling Laws, Carefully | Lil'Log
Embedding parameter count matters for small models. In the small-parameter regime, embedding parameters are a non-negligible fraction of the total and thus counting them or not matters. Pearce & Song (2024) did a thorough analysis along this line. Let’s use to denote model size and compute when embedding is excluded and use to count total parameters.
Scaling Laws, Carefully | Lil'Log
To celebrate his success, he built his family business an elaborate office on the outskirts of Hsinchu. The building borrows from the colonnaded grandeur of Versailles, its stark white facade lending it an air of presidential authority. Inside, cabinets are filled with contemporary art and jeroboams of prized kaoliang, a Taiwanese sorghum liquor.
Welcome to the Luxury City Built by Taiwan’s A.I. Boom - The New York Times