Open AI claimed that they averaged 3 hours of agent time for each problem in the Oct 6 disclosure. Astra costs something like $50 per hour, so it seems like they used around $150 of compute per problem.

Open AI used millions of dollars of compute on the Navier Stokes example.

From the article, "...who have warned that this could dramatically double the 8-12in rise in sea levels already caused by climate change over the last century."

Interestingly, world ocean levels have increased by about 6 to 9 inches in the last 100 years, but the contiguous US has been sinking, so the "relative sea level has risen about 10 - 12 inches" in the lower 48.

See, https://sealevel.nasa.gov/faq/...

Thank you Brain-Fu. I figured that the Trump administration would have some kind of justification for stopping the wind power project, but I had no idea what it could be.

The EFF wrote, "The Senate Commerce Committee is set to consider S. 737, the SCREEN Act, ... will force millions of adult internet users to give up their anonymity, privacy, and security before they access lawful speech." Verifying a person's age can be done without exposing their identity. It requires a trusted third party, such as the government, to issue age certificates. The internet company can then send the user a random challenge (just a random number), and the user can then prove that they possess a valid age certificate stating that they are over 21 without disclosing the certificate, their identity, or their actual age. This software technique is called a Zero-Knowledge Proof.

A million dollars after taxes (if any) and legal fees is a good start. Add more if he has to fight for compensation. It's more than 18 months in jail which is bad enough. The process that led the conviction needs to be reexamined.

The price of running a 60 billion parameter AI based on an open source model is not much more than $0.50 per hour which is around 10K tokens because it can be run on an expensive laptop or desktop. The price to run that model will not go up unless the cost of those 80GB RAM laptops increases. (The laptops are 10x slower than chatGPT when doing text only input and output.)

The U.S. National Helium Reserve was created in 1925. At its peak in 1960s, it held about one billion cubic meters of helium. Worldwide consumption is about 200 million cubic meters per year. But, the US sold most of it off after passing the Helium Privatization Act of 1996. It would have been nice to have it around now.

It's a bit surprising to me that the author of one of the greatest science fiction novels of all time dies and he gets about 15 comments on the Slashdot post. I guess if he had had more than a million book sales, then he would have been better known.

Suppose that Elon wanted to put 1 GW of solar in orbit and used that GW to power GPUs. What would the cost be? Let's assume that SpaceX can get starship to land both stages and SpaceX can put 100 tones in orbit for $20 million in one launch.

All figures below are WAG - wild ass guesses.

The cost of all the engineering research needed put a GW data center in orbit = ???? = $30B ???"

10,000 tons of Solar Cells - cost $500 million - needs 100 launches = $2B

20,000 tons of Radiators - cost $20B - needs 200 launches = $4B

1.5 million GPUs which weighs 2,000 tons - cost $60B - needs 20 launches = $200M

So, the total cost is around 30 + 3 + 20 + 4 + 60 = $120 Billion (approximately). If you build an AI data center on Earth, most of the cost is the cost of the GPUs. In space, the GPUs are about half the cost. I did not include the cost of communications bandwidth. That might be hard and expensive. On the other hand, "Never underestimate the bandwidth of a station wagon full of tapes [or hard drives] hurtling down the highway." (Replace station wagon with a starship.)

I looked through some AI benchmarks and it seems to me that Open AI, Google, and Anthropic are dominating. The Chinese models seem to be the best open source models and they are not far behind the other models. I think that X's Grok is not far behind also.

IMHO, the "Best" benchmark which has images as part of its input is MMMU (https://mmmu-benchmark.github.io).

Top 5 on MMMU: GPT4o, Claude 3.5 Sonnet, Gemini 1.5 Pro, Claude 3 Opus, and Gemini 1.0 Ultra.

Best Chinese Model is Qwen-2.5-VL-72B which is about 10th.

IMHO, the "Best" text based benchmarks are MMLU and ARC

Top 5 MMLU: GPT-5, GPT-4.1, Claude Opus 4.1, Gemini 2.5 Pro, and Llama 3.1 405B.

Best Chinese Model: DeepSeek-R1

Top 5 ARC: Llama 3.1 405B, Claude 3 Opus, Claude 3.5 / 4.5, AI21 Labs, Meta / Mixed models.

Chinese Model: No data found.

It's hard to know what to put in a science book that would not be revised in 1000 years. You can put any math theorem that has been verified by Lean which includes 95% of undergraduate math. Almost everything else is less certain. I guess that you can put in Newtonian mechanics with the caveat that it is very accurate only for objects with mass between 1 nanogram and one sextillion kilograms moving at less than 0.1% of the speed of light. You could put in many chemical reactions and the periodic table up to atomic number 103(Lawrencium) and maybe add factual information for atoms with atomic number between 104(Rutherfordium) and 113(Nihonium).

It would be interesting to try to figure out which scientific "facts" will still be considered true in 1000 years.

GPT Summary with some added by me.

When Einstein completed his studies at the Swiss Federal Polytechnic in Zürich in 1900, he hoped to enter academia but faced obstacles largely of his own making. He preferred learning independently rather than attending lectures, which gave him a reputation for disregarding authority, and some professors saw him as disrespectful. He also clashed openly with Heinrich Weber, a senior professor who had significant influence over academic hiring. Because of this conflict, Weber refused to write recommendation letters for him—an essential requirement in European academic circles at the time. Without such support, Einstein could not obtain the typical entry-level academic positions, such as research or teaching assistantships. Compounding this, his early doctoral work progressed slowly, and his first dissertation attempt was rejected; he did not receive his Ph.D. until 1905, after already working at the Swiss Patent Office. As a result, he ended up there rather than in a university post. 1905 was the year that he published four famous papers on the Photoelectric effect, Brownian motion, Special Relativity, and E=mc^2.

(I do not have a degree in physics! I took about 7 courses in physics in college.)

LOL :)

The article states "Charleston's sea level rose 13 inches over the past century and faces another four-foot rise by 2100." The current rate of sea level increase is about 4.62 mm (0.182 in)/yr based on satellite data. If sea levels continues to rise at that rate, the total rise by 2100 would be 0.35 meters (13 inches), but the rate has been increasing so we should expect more than 0.35 meters. The IPCC has predicted a global sea level rise between 0.38 meters to 0.77 meters (1.24 to 2.53 feet). (See IPCC Report .) If some time in the future, the Antarctic ice cap melts, that would cause a sea level rise of 58 meters. The melted ice from Greenland would add another 7 meters. (See antarcticglaciers.org.)

'This must be Thursday,' said Arthur to himself, sinking low over his beer. 'I never could get the hang of Thursdays.'