Sentinel minutes for week #51/2024: OpenAI o3 model, a reference class for solar superflares, and a streetlight effect in H5N1
Sentinel status: Greenish 🟢
Top items:
OpenAI announced its new o3 and o3 mini models, which represent a significant advance in AI.
Significant streetlight effect in H5N1; US doesn’t have the majority of human cases in 2024, just the majority of observed cases.
Study builds a reference class of stars like the sun, finds that they produce a solar flare much stronger than the Carrington event about once every hundred years.
Biorisks
H5N1
In the US, H5N1 bird flu continues to spread. More than 64 people in the US have been reported to have H5N1 infections this year.
A confirmed H5N1 infection was reported in a person in Wisconsin who worked on a poultry farm. The Louisiana resident with an H5N1 infection is the first severe case of bird flu in the US and is hospitalized in critical condition; the person is 65 years old and has other medical problems. The person had contact with sick and dead birds in a backyard poultry flock, and the CDC has confirmed that the genetic sequence of the strain causing the patient's infection belongs to an avian lineage.
The California governor has declared a state of emergency over ongoing bird flu infections in dairy herds and risks to dairy workers.
Bird flu continues to spread among commercial and backyard poultry farms in the US. More than 13M poultry have been affected in December alone and is driving up US egg prices to all-time highs. Bird flu detections in poultry flocks usually rise and fall with the fall and spring waterfowl migrations and are usually lower in December than in November, but instead, the numbers of detections and birds affected are both higher this month than last month, suggesting that farm-to-farm spread may be responsible for some portion of current spread.
Cats in Los Angeles County have been infected, as well; some have been exposed to raw milk that likely contained infectious virus, while others may have eaten raw meat from infected animals. Bird flu is also killing wild cougars and has infected a raccoon in Washington state. A total of 26 species of wild mammals are known to have been infected with H5N1 in the US in the past year and a half.
Helen Branswell of STAT News summarizes the H5N1 situation in North America in a 2-minute video.
An article from the Center for Strategic and International Studies argues that the US faces serious risks from H5N1 and recommends the following actions:
"Enhance disease testing and surveillance."
Advance the development of vaccines and therapies."
"Strengthen protections for farm workers."
"Work with global partners."
An article by Amy Maxmen for KFF Health News lays out "How America lost control of the bird flu, setting the stage for another pandemic."
Meanwhile, however, forecasters think that there is a “streetlight effect”, and that the US doesn’t have the majority of human cases, just the majority of detected cases, because it has better surveillance. In wildlife populations, we know H5N1 is spreading widely around the world. Spread to domestic animals is most likely along waterfowl (ducks, geese, swans, etc.) migration routes, the natural host species for avian influenza.
Forecasters in aggregate think that 6% to 40% (mean: 20%) of human H5N1 cases this year happened in the US. Here is some extended commentary:
The US has ~4.23% of the world's population but apparently 100% of H5N1-infected cattle. Is H5N1 in cattle in Mexico or further south? It doesn't seem to be in Canada—they do some surveillance—but Mexico's surveillance is much sparser. I find it somewhat hard to believe that cattle H5N1 hasn't left the US. But anyway, if we say that at least most H5N1 in cattle is in the US, then that means most infections from dairy farms are in the US. And the US seems to have a lot of those. And it seems that dairy workers at infected farms may be more likely to get H5N1 than are poultry workers at infected farms, but I'm not sure.
But what about infections from domestic and wild birds? Infections originating in wild birds seem to be relatively rare and to represent a small proportion of human infections. And it's not ambiguous when poultry flocks are infected. A lot of flocks are infected around the world. And wherever flocks are infected, a few people tend to be infected. Not every flock probably, but here and there.
So when I look at the world's highly pathogenic avian influenza outbreaks in poultry farms over the past year, and I factor in under-reporting in various places (e.g., Africa, South America and some Southeast Asian countries), it would seem likely to me that maybe 10% of bird-origin human infections are in the US. Then maybe add another 15% for all the cattle infections? Could be 5%, could be as high as 50%. Probably not more than that. There are probably a very large number of avian-origin cases we never hear anything about. Bird flu in DRC? We'd probably never hear about it. Anyway, I'll go with 25% because of dairy cattle.
Mpox
Four members of a family in Germany were found to be infected with clade 1b mpox virus; clade 1b is the genetic group of mpox viruses that spreads most easily from person to person. Two children were infected, and their school was closed, as a precaution.
Almost 70,000 mpox cases of all clades have been reported in 20 African countries so far in 2024. Case counts are still increasing in Burundi, most likely because of clade 1b spread.
Other biorisks
The "mystery illness" reported in the Democratic Republic of the Congo is likely caused at least in part by malaria, as most cases have tested positive for malaria. However, not all of the cases might be caused by malaria, as a man in Congo has died of hemorrhagic fever. As the AP reports, "That has led to a working hypothesis that the outbreak is either 'severe malaria on a background of malnutrition' or 'a viral infection that is happening on the background of malaria,' [Africa Center for Disease Control and Prevention official] Ngongo told reporters."
A disease called dinga dinga, which causes flu-like symptoms and shaking, has been affecting small numbers of mostly women and girls in Uganda since early 2023. The disease is usually treatable with antibiotics.
Geopolitics
Middle East
In Syria, Hayat Tahrir al-Sham (HTS) continues to consolidate its rule. US officials met with HTS leader Ahmed al-Sharaa and removed a $10M bounty on his head. HTS appears to have been largely self-funded in recent years from taxes in its territory as it embarked on a pragmatic approach to governance in the areas it controlled and managed a border crossing. It is hence not likely beholden to any country and not likely to lead a puppet regime.
Israel continued to bombard military targets in Syria and to hold territory in a demilitarized buffer zone between the countries.
Donald Trump is “considering a proposal to strike Iran’s nuclear programme,” and so is Israel, now that they can freely operate in Syrian airspace after the fall of the Assad regime. Forecasters think a US strike on Iran’s nuclear program is unlikely, with forecasts ranging from basically 0% to 29%, with arithmetic mean 12%.
Iran is facing an extremely dire energy crisis, “the most severe since the 1979 revolution.” Schools are closed, and industries face days to weeks of power cuts, as scarce natural gas is allocated to households for heating homes rather than to power plants. The situation is likely to cause large economic problems and hence to increase political instability in the country, especially as the crisis comes on the heels of massive failures of Iran’s foreign policy in Syria and elsewhere in the Middle East, through weakening of its proxies Hezbollah and Hamas. Moreover, these failures have occurred on a background of economic hardship caused in part by sanctions, and of disgruntlement with the regime, as has been seen in the “Woman Life Freedom” demonstrations two years ago and in declining mosque attendance, which has caused two thirds of the country’s mosques to close.
An article in Time magazine examines how Israel has been using AI to systematically identify airstrike targets, following up on earlier reporting. “‘During the period in which I served in the target room [between 2010 and 2015], you needed a team of around 20 intelligence officers to work for around 250 days to gather something between 200 to 250 targets,’ Tal Mimran, a lecturer at Hebrew University in Jerusalem and a former legal adviser in the IDF, tells TIME. ‘Today, the AI will do that in a week.’ One intelligence officer tasked with authorizing a strike recalled dedicating roughly 20 seconds to personally confirming a target, which could amount to verifying that the individual in question was male.”
Africa
Following the collapse of the Assad regime in Syria, Russia is attempting to increase its military presence in eastern Libya, which is controlled by the warlord Gen. Khalifa Haftar and his family. Russia is reportedly increasing pressure on Haftar to allow its navy access to the port in Benghazi. The internationally recognized Libyan government, which controls western Libya, is pushing back. Increasing cooperation between Haftar and Russia can be expected to shift the balance of power in Libya towards Haftar and to affect Russia’s projection of power in the Mediterranean and in Africa.
At least 110 people have died in 7 weeks of post-election protests in Mozambique. Protesters alleged that the presidential election was rigged in favour of the long-entrenched Frelimo party, which has been in power since the country gained independence in 1975. Protests escalated following the killing of two opposition officials.
Europe
The leader of Russia's nuclear, biological and chemical forces was killed in a bomb blast in Moscow claimed by Ukraine. He had been accused of using chemical weapons against Ukrainian troops.
UK troops could be sent to Ukraine to train soldiers there, according to Defence Secretary John Healey. This could ultimately pave the way for UK troops to monitor a ceasefire or deter Russia if it achieves a sudden breakthrough.
Germany's Chancellor Olaf Scholz lost a no-confidence vote. There will likely be snap elections on February 23. Scholz' party, the Social Democrats, is trailing the Christian Democratic Union in polls.
The European Commission has awarded €10.6 billion in contracts for a secure communications satellite network for use by EU governments, EU public agencies, and commercial services. The network, known as IRIS, will have more than 280 satellites.
In Germany, a man drove a car into a crowded Christmas market, killing at least 5 people and injuring more than 200. The perpetrator was a 50-year-old Saudi doctor who was a permanent resident in Germany and reportedly supported the AfD and anti-Islam politicians.
Switzerland plans to spend £200m to upgrade nuclear shelters.
The US and the UK imposed sanctions on several Georgian officials.
The US will announce a new military aid package for Ukraine in the coming days. This will include air defense interceptors and artillery munitions.
Asia Pacific
China’s short-term bond yields fell below 1% for the first time since 2009, likely reflecting weak market expectations for the Chinese economy. Indeed, the Chinese economy is facing a rough year ahead.
A newly released report by the US Department of Defense on security and military developments in China through early 2024 highlights China's ramping up of activities around Taiwan, Taiwan's focus on the development of both asymmetric and conventional defense capabilities, and China's increasing activity in the South China Sea.
In Taiwan, lawmakers brawled over three bills supported by the Nationalist Party that President Lai Ching-te's party, the Democratic Progressive Party, says would weaken Taiwan's democratic system.
President Biden has approved $571m in defense support for Taiwan.
Next month, the Bank of England governor, Andrew Bailey, will visit China with British finance minister Rachel Reeves.
The Americas
The US Senate has passed a government funding bill to avert a US government shutdown.
Guyana is cooperating more with the US amidst conflict with Venezuela and ExxonMobil operating in contested waters, reports the Venezuela press.
The Russian air force flew nuclear-capable bombers in neutral waters near Alaska.
Two Americans of Chinese descent were running a secret Chinese “police station” in NYC. “The outpost, which occupied an entire floor above a ramen stall, did provide basic services such as renewing Chinese citizens’ driver licenses, but it also helped Beijing to identify pro-democracy activists living in the US, say federal authorities.
The US Supreme Court will hear TikTok’s case on their pending US ban.
Artificial Intelligence
o3
OpenAI announced its o3 model, along with o3-mini. They plan to deploy the models in early 2025. o3 scores a breakthrough 75.7% on the ARC-AGI Semi-Private Evaluation benchmark, which is now beginning to get saturated. However, the model still fails on some tasks that are easy for humans.
o3 also makes significant advances on benchmarks for math, science, and coding, scoring 2727 on Codeforces, equivalent to the 175th best human competitive coder on the planet.
A member of OpenAI’s preparedness team has said that o3 is AGI. The post was then later deleted. Most of our forecasters disagree with this assessment due to its cost (and some use more stringent definitions of AGI than OpenAI themselves do).
o3 represents a significant advancement of the frontier of AI.
Forecasters expect the cost of using a roughly o3-equivalent model to fall about one order of magnitude (aggregate range 0.2 to 1.8 OOMs) between now and mid 2025.
Alignment faking
Claude—the name for Anthropic’s line of LLMs—normally refuses harmful prompts. Researchers at Anthropic and Redwood Research told it they were instead training it to comply with harmful prompts and set up a scenario in which Claude thought it was monitored for normal users but unmonitored for enterprise users.
Authors present results as “when monitored it faked alignment 12% of the time.” One of the authors wrote: "Thank god this model is aligned, because if not this would be scary" is imo basically the correct takeaway from our work. The values in fact aren't scary! The scary thing is that the model protects its values from our attempts to change them.”
A commentator instead characterized this behavior more specifically as “gradient hacking”.
Conversely, other commentators characterized this either as a catch-22—bad if they comply with harmful user requests, bad if they don’t—or as philosophically confused.
Other AI developments
Alec Radford, a longtime researcher at OpenAI, left the company. He was the lead author on OpenAI’s original GPT paper.
Italy fined OpenAI $15.58m over breaching privacy rules.
Ukrainian drone footage from the war could be useful training data for training AI models to make decisions on the battlefield.
Elon Musk has said that xAI will offer free access to Grok-2.
People keep falling for fake, AI-generated videos that are claimed to be of drones flying over New Jersey in the US.
The German data protection authority in Bavaria has ordered Sam Altman’s biometric ID project “World” (formerly Worldcoin) to delete data.
Natural Disasters
A new study of over 2500 stars like our sun has found that such stars probably release a superflare on average about once every century. The energies released in these flares are between 100 and 10,000 times as much as the energy released in the largest solar flare ever measured, which caused the Carrington event. It is not known whether such superflares also trigger coronal mass ejections, which would pose a danger to Earth if hurled in its direction.
We think the above reflects positively for our previous estimate based on Laplace’s law and the historical record, and negatively for research for Open Philanthropy done by David Roodman.
The New York Times reports that in the US, insurance companies have stopped insuring more than 1.9 million homes since 2018 because of increasing climate-related risks. As the article notes, "Without insurance, you can’t get a mortgage; without a mortgage, most Americans can’t buy a home. Communities that are deemed too dangerous to insure face the risk of falling property values, which means less tax revenue for schools, police and other basic services. As insurers pull back, they can destabilize the communities left behind." Over time, uninsured losses will increase poverty and political instability.