HNHacker News
TopNewBestAskShowJobs

treebeard901

810 karma · joined October 16, 2015

submissionscomments
treebeard901··on Generate fonts where every LLM token is the same width
Yeah that is one draw back they mention in the paper. So much of the full test stack is trained on and dependent on English.

A few months ago I went down this same rabbit hole trying to determine if using Chinese has any inherent advantages due to the way more context is encoded into Chinese characters versus what you find in English.

The total token size may not take into account any benefits gained by Chinese characters potentially encoding more context.

I have been meaning to see if any Chinese native tokenizers have been created to account for this feature of the language.

Unfortunately, many Chinese and Japanese characters can have many different meanings depending on how they are used. As a result they do not break down as well to be tokenized as a language like English.

Maybe some kind of hybrid intermediate language would be most optimal.

treebeard901··on If you don't have the factories, you lose the expertise
What is the economic war being waged on "the West"?

The U.S. suffers from a geopolitical bubble which operates like a financial bubble. Inside the bubble, global nations rebuilding and trading without the U.S. reserve currency is perceived as war. When it reality it is just the natural course of countries and economies after WW2.

treebeard901··on Pentagon says overreliance on AI contributed to missile strike on Iran school
Its difficult to believe they were not aware it was a school before ordering the strikes.
treebeard901··on U.S. Strike on Iranian School May Have Been War Crime, U.N. Report Says
So many unanswered questions from this. Right after the narrative was that some form of AI was used to determine the targets during the initial campaign. Presumably they captured a lot of location data from IRGC and other Govt leaders to find places where they cluster at specific times and places. There are very few schools the Iranian elite would send their kids to and so there was heavy concentration at this specific place and time. They had so many targets in the opening days that maybe they only used certain graphing algorithims and took the ones with the highest centrality scores and the most common repeating place and times for the targets they had in mind. But then you have to wonder if some human somewhere didn't cross reference this supposed AI output to determine what specifically they were about to bomb.

That question should be answered. Did human review of this target reveal it was a school? If so, did someone in the Govt approve the strike knowing this?

Was it too tempting of a target given who was likely in the same place at the same time (dropping off or picking up their kids from school) ? Did they use phone location data and since parents buy phones for their kids, it looked like a lot of IRGC and other state officials were at that location, when infact it was their children and teachers?

Even more disturbing, if they knew it was a school and still chose to strike, was the intent to wipe out the next generation of Iranian leadership?

Just so many unanswered questions with no good answers.

Of course it was a war crime.

treebeard901··on Training a 4B model to produce 81% faster query plans than Postgres
If the LLM is writing the SQL it might as well build the query plan too.
treebeard901··on How good are frontier models at physics?
> To our temporary relief and encouragement, it has not managed to fully resolve even one ...

Why would that be encouraging if the purpose of their chosen profession is to advance physics? This is sort of like mathematics being solved by AI with human researchers complaining that they are losing potential awards. It's an entirely self serving way to be.

treebeard901··on The engineering behind the US Strategic Petroleum Reserve
I always thought the smartest thing to do since countries chose to accept fiat currency in exchange for oil was to create as many patroleum reserves as possible and print as much money to buy as much oil to store for later use as they would sell.
treebeard901··on Salesforce Global Outage
All the prompt engineers clogging up San Francisco for dreamforce need to get back to the office.
treebeard901··on Autism is genetic – and why this information matters now
It's an evolutionary trait that is in response to increasing use of electronics and computers. As being good with technology is a larger part of survival of the fittest, the autism traits will only increase along with it.
treebeard901··on A single firm is behind OpenAI, Anthropic, and Meta hacking scandals
Its so manipulative for IPO reasons too. The fact is that all of these systems are still just stastically predicting the next token. Spending trillions on data centers and more training data hasnt changed the fundamental reality that it is still just predicting the next token.

That is not worth the investments being made.

treebeard901··on Steam Frame starts at $1059
Just duct tape your smart phone to your face for free
treebeard901··on Oil Executives Say the Great Fuel Crisis Is Here
It's all fracked up
treebeard901··on Physicists can't agree on how the Universe works
> 68% of the physicists surveyed said the Big Bang does not necessarily represent the beginning of time

Kind of an interesting rabbit hole with this one...

treebeard901··on Nvidia is the central bank of AI
The Japanese economy and yen carry trade is a close second.Rising oul prices and reduced output due to the conflicts in the Middle East could filter through to increasing yields on Japanese debt. In turn, the yen interventions have to continue to keep it lower than 160, which seems to be the psychological barrier for the yen carry trade.
treebeard901··on Music Theory for the 21st-Century Classroom
The history around how 440hz came to be the standard is pretty interesting in itself.
treebeard901··on OpenAI considers slowing advanced AI development, Sam Altman tells employees
Hugging Face wasnt really an escaped model that presents a kind of future threat. The task the model was given lead it to discover certain admin credentials that were left publicly in Github. All it did was use that to complete the task.

With other countries and open source, not to mention all the other companies, Open AI choosing to pause development wont change the course. Only OpenAI will lose.

Unless the secret is that no amount of compute or data can progress things any further because we have already reached the limit for what predicting the next token can do. Most of the advancements lately have been on the agentic side of things.

The massive amount of health, science and math problems being solved by AI has largely not materialized yet. Is this a sign the model capability has been over sold?

There are economic consequences to this as well because the data center build out has been a kind of circular financing. This has caused all kinds of macro economic impacts from increased prices to even the Treasury market itself having to compete with debt issued for the build out. No one knows what will happen once the data center build out stops, or if the yen carry trade continues unwinding, or if Treasury yields have to keep rising to compete with debt issued by the "Mag 7".

Once they stop spending in the current circular arrangement, then the profits that exist on paper disappear, and the bubble could pop.

So really for every reason, OpenAI cant really slow development. Unless there is a kind of Moores law for what predicting the next token can achieve.

treebeard901··on Show HN: GET Together – A social network where you don't need POST to Post
Server capacity reached. Upgrade now to continue your session!
treebeard901··on Muse Spark 1.3
Or placed in an asylum next to the people who designed XML
treebeard901··on Mamdani Bans AI in NYC Schools
If AI has a major impact on the job market as some believe, then it would be interesting to see a society where AI related profits are used to provide free education for everyone who wants it.

At some point they will begin running out of human created data for model training and schools with strict AI policies will be the best source of non AI data going forward.

Thats kind of a symbiotic relationship that makes sense for students and AI companies.

Once AI output begins replacing human output and the models begin training on it, there may be increasingly few places left for data that isnt likely to lead to a model collapse.

treebeard901··on Omarchy: Any User Process Can Escalate to Root
I was just adding more context than your post included
treebeard901··on 'Talk to Mom': Burning Man attendees use this phone to call home (2024)
San Francisco should bring back the original Burning Man at Bakers Beach.
treebeard901··on Omarchy: Any User Process Can Escalate to Root
DHH was on the Lex Fridman podcast talking about this recent release of Omarchy in that most of it is "vibe coded". It is mostly just a bash script to configure Linux, but his approach is interesting.

AI as a core part of the OS that can just change or add anything you want. Linux is great for this because it has access to the source code for everything.

He said he didnt review the code line by line, just looked at the shape of it. Whatever that means.

treebeard901··on 9th Circuit sides with states in Kalshi gambling fight
Kalshi hired Donald Trump Jr as a strategic advisor with a salary reported to be around $300,000 per year. Recently there was a meeting between him and various State Attorneys General, where the message was back off prediction markets. Not too long after that other areas of the Trump Admin sent a letter basically saying the same thing.

And so the circuit courts coming to very different conclusions from the 9th circuit represents the inherent political corruption involved in the court system.

Circuit splits get resolved at the Supreme Court and I would make a prediction that it isnt a coin flip, probably more of a safe bet.

treebeard901··on The turbulent AI era is here
Every person says this even in notes
treebeard901··on AI Chip Architectures
Recently I started thinking about an analog optical device capable of matrix multiplication. After some online research, the Mach-Zehnder interferometer seems to have been adapted by a German company into a full addon optical based matrix multiplication system.

It is a very interesting concept and the company Q.Ant has delivered working products and software. Apparently it can greatly speed up certain training applications and future designs will solve some of the current problems.

The energy and performance advantages seem to be worth the R&D. If some form of optical memory to replace DRAM in the chain could be designed, that would solve several conversion costs between optical to digital.

Either way it is pretty cool to do math this way, basically just using light.

treebeard901··on Xiaomi: New CPU matches Apple cores single threaded, much faster multithreaded
Using the above post example of a delay when typing in a terminal, the path that text takes from the keyboard to the screen is much more complicated. Some of it comes from security changes. Back in the day everything had direct memory access. And the memory layout wasnt randomized by things like ASLR. Not that this specifically should impact terminal input but it becomes a factor with memory access more generally. Now you have to deal with user mode and other security barriers. Drivers can still get direct hardware access but then as has been mentioned you have the bloated software layer with stuff like electron. Increased concurrency has a latency cost as well and we have many more things (usually unnecessarily) running today.

Im sure there are many more things that add up to focus more on throughout over latency as mentioned above.

treebeard901··on Rethinking Database Programming
The main arguments presented are that databases do not support modern types.. And that this system replaces tested authentication systems by emailing a UUID in plain text?

It does mention UInt64 which is not a modern type and as far as I know is supported by every database.

It also compiles to SQL but it isnt clear where the advantage comes from other than using a different syntax to do things.

treebeard901··on Norway Should Buy OpenAI
If AI has a negative impact on employment as some think then AI should be required to fund free college for everyone that wants it. Then the schools can have a strict rule that only limited AI can be used during your education.

Then set up a pipeline between the human generated content from the colleges back into future AI models. Everyone gets an education and there is ongoing non AI training data to advance AI without any model collapse.

treebeard901··on People are worried about America's solvency
If the United States tried to make various geopolitical moves to sabotage its reserve currency and ultimately become insolvent as a result of a self inflicted debt crisis then I don't see how the current agent orange situation could be topped.

There are so many plumbing issues with the financial system happening at once that people really do not have a clear picture of what is happening in totality.

Start with Japan... Still the largest foreign creditor. However they are facing a situation where the yen carry trade that began after the plaza accords in the 1980s is unwinding. The recent Treasury intervention and the associated expansion of the repo market so they do not sell treasuries directly reveals a hidden structural weakness. This leads to the AI boom because cheap Japanese debt was recycled into larger yields in the U.S. And it's different this time because companies like Microsoft have what is rated as safer debt than the US Treasury itself. The yield on financing part of the AI boom has been much better than yields back home in Japan. So even Japanese institutional investors may start seeing domestic debt as more attractive as the carry trade unwinds and domestic Japanese debt starts paying more. This ties back into the recent treasury intervention and offering a way to unload U.S. debt without impacting the markets directly.

Which brings us to an ongoing structural change that will remove the largest foreign investor in U.S. debt. Which as we all know, is approaching $40 trillion with over $30 trillion of that being held as public debt. Meanwhile, the budget deficit is continuing to rise causing even more issuance of debt. After the 2008 crash, and over the next decade, the U.S. was able to issue debt basically for free and it issued a ton of it. But that debt matures and when it matures, it has to largely be reissued at current rates. There was no free lunch. With the Japanese largely pulling back and at best not selling the debt they have now to defend their currency or to chase domestic yields in Japan, that goes to the UK as the second largest foreign investor. And the UK clearly can't absorb what the US has to issue and refinance. China has been reducing its holdings because as the BRICS bank, etc, replace their need for dollars, they will begin using it in more creative ways.

Some speculation is that the U.S. intervention into the yen carry trade is being challenged by foreign actors, which did not happen during past interventions. Monitoring the daily spread, and if the yen continues climbing over 160 then this will be the first time the U.S. intervention failed to work. To prevent selling existing Treasuries to defend the currency, the Japanese investors will unload it in the repo market.

But this brings us to a new structural problem in response to the insane amount of debt being issued and refinanced and this is called the weighted average maturity. Recently, the U.S. Treasury implied that they are going to begin moving more issuance to shorter term debt because those markets are currently still much more liquid than the longer term markets that had issues revealed by the failure of Silicon Valley Bank. When SVB needed to cover deposits, they found the 20 and 30 year Treasury market was not as liquid as they had hoped. This same fact likely pushed the Treasury to have Japanese investors move towards the repo market instead.

But this brings a new problem, because there is only so much appetite for increasingly taking long term debt and reissuance and moving it towards the front of the weighted average. Currently around 6 years, and the goal is to move it between 2 and 4. Imagine the liquidity impact on moving large amounts of new debt issued and refinanced debt from the 30yr to more short term funding.

Next up, there is a need to have the Federal Reserve unwind its balance sheet even more quickly at the same time all of the above is happening. But it is the Feds balance sheet and forms of QE that have propped up the repo market itself under times of stress. So after Japan, the UK and China all have reasons to continue cutting back, and after the short term funding gets eaten up, the yields will have to rise dramatically to continue attracting investors for even the short term. The immediate response to this will first be to stop issuing the 20 year and reduce the 30 year. They may debate the 50 year but odds are if the 30 has a weak showing that there would be even less appetite for the 50.

But with the unwinding yen carry trade, we have to consider the impact on funding the AI boom going forward. Clearly this has been on the mind of nVidia recently, with the move to create a form of mortgage backed securities for GPU capacity. The largest tech companies have something like $2-3 trillion in shell companies set up specifically for the data center buildouts. The rest of the revenue seems to be going in a circular motion among several large companies. They know as soon as they stop funding the AI buildout, or are forced to by the unwinding of market forces, that the party will end. This is the very reason for the move from nVidia recently. And if mortgage backed securities don't make you feel better about it, they have also compared it to how airlines trade airplanes as assets. Neither is a comparison that they should want to make for historical reasons.

Now in the past, the U.S. allies would see the value in the financial alliance and would ride to the rescue of something like a 50 year bond to lock up these bubbles for a long time. After so many years of tariffs and other forms of coercion, even the recent yen internvention broke norms between the U.S. Treasury and the European Central Bank. Maybe in the past the Saudis, UAE, etc would have jumped in to invest, but clearly those days are over too... With massive implications for the petrodollar. Which itself is another major factor in the reserve currency system and is clearly under strain from the Iran War and other geopolitical factors.

The U.S. response to a debt crisis won't be to default, it will do what it always has done, and try to inflate itself out of the situation. It's just that they can't really take this path either. The more the Federal Reserve becomes the lender of last resort and expands its balance sheet to backstop moving all this to the short term repo market, the end result will be extreme inflation. The yields will begin rising on short term debt faster than we have been seeing on long term debt. Before long, it will be early 1980s style interest rates, and if they don't find a solution, this process will be self repeating... The $40 trillion, with roughly $1.2 trillion in deficits every year, and past reissuance, all into short term funding markets, where the only buyer capable of absorbing it all is the Federal Reserve, there is just no real way out of it.

Even all the structural problems with mortgage backed securities haven't been fixed since 2008. If the AI bubble pops, the energy prices keep increasing, the yen carry trade continues to unwind, any South East Asian conflict with North and South Korea, China/Japan/Taiwan, etc... Or with the Ukraine war ending poorly with Russia moving further into the baltic states...

Maybe hope that a guy who bankrupted several casinos can figure out to do when the largest casino in history runs out of money? I don't know.

EDIT: Sorry for the wall of text. Had a lot to say.

treebeard901··on Judge relying wholly on AI in order is covered by judicial immunity, court rules
All officers of the court have various forms of immunity. All of the rights, rules, and statutes they are required by law to uphold only goes as far as how much of that in group someone is, or how much money someone has to uphold their rights when infringed upon.

In practice the entire system has become political, corrupted or both. As a result, the only rights most people have are those where officers of a court are only limited by where their immunity ends. They can and do violate the law and ignore large parts of it if it suits them.

Assuming someone has the resources to even get any oversight at all, in most cases those involved will investigate each other and decide they have done nothing wrong. Even in the unlikely event that someone proves an illegal act, they will just claim immunity.

Page 1 of 14Next →