| Not researched |
Factual |
8 Aug 2023 |
Anthropic's Long Term Benefit Trust is a much narrower body that makes decisions for Anthropic.
Factual · 8 Aug 2023 |
Their own work or organisation |
| Not researched |
Factual |
8 Aug 2023 |
Convolutional neural networks build in translational symmetry.
Factual · 8 Aug 2023 |
AI capabilities |
| Mostly true |
Factual |
8 Aug 2023 |
Anthropic published a blog post on security stating it has a two-key system for access to model weights.
Factual · 8 Aug 2023 |
Their own work or organisation |
| Not researched |
Factual |
8 Aug 2023 |
The US is substantially ahead of China in AI, though China is trying hard to catch up.
Factual · 8 Aug 2023 |
AI policy and governance |
| Own records only |
Factual |
8 Aug 2023 |
Anthropic has worked to ensure the data center used for its models is located in the United States.
Factual · 8 Aug 2023 |
AI compute, hardware and data |
| Not researched |
Factual |
8 Aug 2023 |
Dario Amodei's "big blob of compute" document has not been made public.
Factual · 8 Aug 2023 |
Their own work or organisation |
| Not researched |
Factual |
8 Aug 2023 |
One way to phrase it, Amodei said, is that current AI models are maybe two to three orders of magnitude smaller than the human brain, when compared by number of synapses.
Factual · 8 Aug 2023 |
AI capabilities |
| Mostly true |
Factual |
8 Aug 2023 |
As of August 2023, mechanistic interpretability is not nearly capable enough to verify that a model is aligned.
Factual · 8 Aug 2023 |
AI safety and alignment |
| Not researched |
Factual |
8 Aug 2023 |
Over the six months before August 2023, the 2023 generation of AI models, compared with the 2022 generation, started to learn math and programming.
Factual · 8 Aug 2023 |
AI capabilities |
| Unresolvable |
Factual |
8 Aug 2023 |
AI scaling laws are sometimes predictable to several significant figures, which is not seen outside of physics.
Factual · 8 Aug 2023 |
AI capabilities |
| Unresolvable |
Factual |
8 Aug 2023 |
As of 2023, many firms rushing to deploy AI systems are not using them in anywhere near the most efficient way possible.
Factual · 8 Aug 2023 |
AI, the economy and work |
| Mostly true |
Factual |
8 Aug 2023 |
Existing alignment methods succeed in aligning AI systems for today's tasks.
Factual · 8 Aug 2023 |
AI safety and alignment |
| Unresolvable |
Factual |
8 Aug 2023 |
Current AI models may be superhuman or close to superhuman at constrained writing, such as writing a page without the letter E.
Factual · 8 Aug 2023 |
AI capabilities |
| Not researched |
Factual |
8 Aug 2023 |
Scaling laws in terms of loss are already starting to bend, as seen in published model cards from multiple companies.
Factual · 8 Aug 2023 |
AI compute, hardware and data |
| Not researched |
Factual |
8 Aug 2023 |
Anthropic did not cause the big acceleration in AI that happened in late 2022 and early 2023.
Factual · 8 Aug 2023 |
Their own work or organisation |
| Own records only |
Factual |
8 Aug 2023 |
Dario Amodei has pushed for better cybersecurity at every organisation where he has worked.
Factual · 8 Aug 2023 |
Their own work or organisation |
| Resolved true |
Factual |
8 Aug 2023 |
Amodei said Anthropic "kind of" already does reinforcement learning, through RLHF.
Factual · 8 Aug 2023 |
AI capabilities |
| Own records only |
Factual |
8 Aug 2023 |
Anthropic's current security measures are not sufficient against a super determined state-level actor.
Factual · 8 Aug 2023 |
Their own work or organisation |
| Not researched |
Factual |
8 Aug 2023 |
It is possible to predict above chance from an MRI scan whether a person is a psychopath.
Factual · 8 Aug 2023 |
Science and research methods |
| Not researched |
Factual |
8 Aug 2023 |
In very broad terms, a guess that current frontier models such as GPT-4 and Claude 2 cost on the order of one hundred million dollars is not wrong as to order of magnitude.
Factual · 8 Aug 2023 |
AI compute, hardware and data |
| Mostly true |
Factual |
8 Aug 2023 |
Anthropic's initial constitution for its models drew on broadly agreed sources such as the UN Declaration of Human Rights and parts of Apple's Terms of Service.
Factual · 8 Aug 2023 |
AI safety and alignment |
| Own records only |
Factual |
8 Aug 2023 |
Dario Amodei formed his view on scaling gradually between 2014 and 2017.
Factual · 8 Aug 2023 |
Their own work or organisation |
| Not researched |
Factual |
8 Aug 2023 |
The scary biology information current models can be made to say is often information that could be found by Googling.
Factual · 8 Aug 2023 |
Biosecurity |
| Unresolvable |
Factual |
8 Aug 2023 |
Interpretability work on language models, such as induction heads, suggests much of the cognitive machinery needed for active agents is already present in base language models.
Factual · 8 Aug 2023 |
AI capabilities |
| Unresolvable |
Factual |
8 Aug 2023 |
AI language models learn to speak fluent Base64, which humans do not typically learn.
Factual · 8 Aug 2023 |
AI capabilities |
| Own records only |
Factual |
8 Aug 2023 |
Anthropic is focusing heavily on cybersecurity, including working with its cloud providers.
Factual · 8 Aug 2023 |
Their own work or organisation |
| Not researched |
Factual |
8 Aug 2023 |
Anthropic works with cloud providers for its data centers.
Factual · 8 Aug 2023 |
Their own work or organisation |
| Own records only |
Factual |
8 Aug 2023 |
Every traditional investor who invests in Anthropic looks at its unusual governance body.
Factual · 8 Aug 2023 |
Their own work or organisation |
| Not researched |
Factual |
8 Aug 2023 |
Email accounts of fairly high-ranking US government officials were recently hacked via Microsoft, which provided the accounts.
Factual · 8 Aug 2023 |
Other global risks |
| Not researched |
Factual |
8 Aug 2023 |
As of 2023, AI models are only beginning to prove relatively simple mathematical theorems.
Factual · 8 Aug 2023 |
AI capabilities |
| Not researched |
Factual |
2023 |
Dario Amodei testified in the US Senate that AI might pose serious bio risks within two to three years.
Factual · 2023 |
Biosecurity |
| Unresolvable |
Factual |
13 Oct 2022 |
Amodei said he does not believe US export controls were ever designed to prevent China from obtaining a few tens of thousands of chips.
Factual · 13 Oct 2022 |
AI policy and governance |
| Not researched |
Factual |
13 Oct 2022 |
Nvidia H20 chips are less efficient for training and more efficient for sampling, and are still allowed under US export controls.
Factual · 13 Oct 2022 |
AI policy and governance |
| Not researched |
Factual |
13 Oct 2022 |
Producing DeepSeek's R1 from V3 was probably very cheap.
Factual · 13 Oct 2022 |
Economics and markets |
| Unresolvable |
Factual |
13 Oct 2022 |
DeepSeek, a Chinese company, was the first to demonstrate the expected AI cost reductions, which had never before happened with a Chinese company.
Factual · 13 Oct 2022 |
Economics and markets |
| Mostly true |
Factual |
13 Oct 2022 |
Because reasoning-focused RL is new, the amount all AI developers spend on the RL training stage is still small, so the field is early on its scaling curve.
Factual · 13 Oct 2022 |
AI compute, hardware and data |
| Resolved true |
Factual |
13 Oct 2022 |
The current rate of algorithmic curve shift is maybe about 4x per year.
Factual · 13 Oct 2022 |
AI compute, hardware and data |
| Own records only |
Factual |
13 Oct 2022 |
DeepSeek-V3 was the real innovation and should have made people take notice a month before this statement, and the speaker says that "we" did take notice of it.
Factual · 13 Oct 2022 |
AI capabilities |
| Unresolvable |
Factual |
13 Oct 2022 |
Anthropic finds Claude 3.5 Sonnet remains much better than DeepSeek-V3 on some key tasks such as real-world coding.
Factual · 13 Oct 2022 |
AI capabilities |
| Mixed |
Factual |
13 Oct 2022 |
DeepSeek was not substantially more resource-constrained than US AI companies, and export controls were not the main factor causing its innovation.
Factual · 13 Oct 2022 |
AI policy and governance |
| Mostly true |
Factual |
13 Oct 2022 |
Dario Amodei and his co-founders were among the first to document, while at OpenAI, that scaling up AI training smoothly improves results across a range of cognitive tasks.
Factual · 13 Oct 2022 |
Their own work or organisation |
| Mostly true |
Factual |
13 Oct 2022 |
Improvements in training cost efficiency typically lead AI companies to spend more, not less, on training models.
Factual · 13 Oct 2022 |
Economics and markets |
| Not researched |
Factual |
13 Oct 2022 |
DeepSeek's team achieved V3's results through innovations mostly focused on engineering efficiency, including Key-Value cache management and mixture of experts.
Factual · 13 Oct 2022 |
AI compute, hardware and data |
| Resolved true |
Factual |
13 Oct 2022 |
Amodei said he thinks a fair statement is that DeepSeek produced a model close in performance to US models 7-10 months older, at a good deal less cost but not near the ratios people have suggested.
Factual · 13 Oct 2022 |
Economics and markets |
| Not researched |
Factual |
13 Oct 2022 |
DeepSeek's total company spend is not vastly different from that of US AI labs.
Factual · 13 Oct 2022 |
Economics and markets |
| Not researched |
Factual |
13 Oct 2022 |
Sonnet's training was conducted 9-12 months ago, DeepSeek's model was trained in November/December, and Sonnet remains notably ahead in many internal and external evals.
Factual · 13 Oct 2022 |
AI capabilities |
| Not researched |
Factual |
13 Oct 2022 |
Claude 3.5 Sonnet is a mid-sized model that cost a few tens of millions of dollars to train.
Factual · 13 Oct 2022 |
Economics and markets |
| Own records only |
Factual |
21 Jul 2017 |
Around 2015, Dario Amodei often argued that a relatively peaceful geopolitical period made it a good time for AGI to be built.
Factual · 21 Jul 2017 |
AI timelines and forecasting |
| Resolved true |
Factual |
21 Jul 2017 |
OpenAI and DeepMind both have a focus on safety and both have safety teams.
Factual · 21 Jul 2017 |
AI safety and alignment |
| Mostly true |
Factual |
21 Jul 2017 |
Getting started in machine learning probably requires about a thousand dollars to buy a GPU.
Factual · 21 Jul 2017 |
Field building and careers |
| Mostly true |
Factual |
21 Jul 2017 |
There was a lot of political instability in the western world in the year before July 2017.
Factual · 21 Jul 2017 |
Philosophy, ethics and society |
| Own records only |
Factual |
21 Jul 2017 |
OpenAI does not lower its machine learning hiring bar for people who want to work on safety.
Factual · 21 Jul 2017 |
Their own work or organisation |
| Unresolvable |
Factual |
21 Jul 2017 |
Vicarious does not think about AI safety, to Amodei's knowledge.
Factual · 21 Jul 2017 |
AI safety and alignment |
| Not researched |
Factual |
21 Jul 2017 |
Dario Amodei was not present when OpenAI was founded or its name chosen, and did not pick the name.
Factual · 21 Jul 2017 |
Their own work or organisation |
| Not researched |
Factual |
21 Jul 2017 |
OpenAI's safety team consisted of Dario Amodei and Paul Christiano as of mid-2017.
Factual · 21 Jul 2017 |
Their own work or organisation |
| Own records only |
Factual |
21 Jul 2017 |
Amodei said that he and OpenAI focus a lot on reinforcement learning because it, and extended versions of it, seems a better fit for what intelligent agents do in general.
Factual · 21 Jul 2017 |
Their own work or organisation |
| Not researched |
Factual |
21 Jul 2017 |
Early expert systems based purely on logical reasoning proved very brittle because they could not represent high-dimensional space.
Factual · 21 Jul 2017 |
AI capabilities |
| Not researched |
Factual |
21 Jul 2017 |
The principles on which a visual neural net is built are simple: backpropagation and alternating linear and non-linear components.
Factual · 21 Jul 2017 |
AI capabilities |
| Own records only |
Factual |
21 Jul 2017 |
Amodei said he thinks that, broadly, most people at OpenAI are worried about the issues under discussion or at least think they are worth thinking about, and he set this apart from the staff actively doing technical work on them.
Factual · 21 Jul 2017 |
Their own work or organisation |
| Unresolvable |
Factual |
21 Jul 2017 |
Dario Amodei mainly works on safety at OpenAI.
Factual · 21 Jul 2017 |
AI safety and alignment |
| Not researched |
Factual |
21 Jul 2017 |
On some Atari games that take about 10 million timesteps to learn, a human usually has to give feedback only a few thousand times.
Factual · 21 Jul 2017 |
AI capabilities |
| Not researched |
Factual |
21 Jul 2017 |
OpenAI as an institution holds that working on AI safety requires being at the forefront of AI.
Factual · 21 Jul 2017 |
AI safety and alignment |
| Own records only |
Factual |
21 Jul 2017 |
Dario Amodei has not been closely involved in OpenAI's rapid hiring and growth, except on the safety side.
Factual · 21 Jul 2017 |
Their own work or organisation |
| Not researched |
Factual |
21 Jul 2017 |
Not many people were yet working on technical AI safety.
Factual · 21 Jul 2017 |
Field building and careers |
| Not researched |
Factual |
21 Jul 2017 |
As of 2017, OpenAI is hiring talented machine-learning people who care about AI safety.
Factual · 21 Jul 2017 |
Field building and careers |
| Mixed |
Factual |
21 Jul 2017 |
For almost a decade before 2017, self-driving cars have existed that are 99.9% safe.
Factual · 21 Jul 2017 |
AI capabilities |
| Unresolvable |
Factual |
21 Jul 2017 |
The Concrete Problems in AI Safety paper did not discuss the transparency of neural nets.
Factual · 21 Jul 2017 |
AI safety and alignment |
| Not researched |
Factual |
21 Jul 2017 |
Researchers trained a simulated robot to do a backflip using human feedback over a few hundred timesteps, with about 30 minutes of human clicking.
Factual · 21 Jul 2017 |
AI capabilities |
| Not researched |
Factual |
21 Jul 2017 |
OpenAI had about 55 staff in mid-2017.
Factual · 21 Jul 2017 |
Their own work or organisation |
| Not researched |
Factual |
21 Jul 2017 |
OpenAI is a nonprofit and does not sell products.
Factual · 21 Jul 2017 |
Their own work or organisation |
| Not researched |
Factual |
21 Jul 2017 |
As of 2017, OpenAI's funding came only from donations.
Factual · 21 Jul 2017 |
Their own work or organisation |
| Not researched |
Factual |
21 Jul 2017 |
Amodei said that his team's paper replaces the fixed reward function with a neural-net-based model of the human's reward.
Factual · 21 Jul 2017 |
AI capabilities |
| Not researched |
Factual |
21 Jul 2017 |
A researcher named Stuart, at a group that sends interns to OpenAI, has been thinking about AI safety for a while.
Factual · 21 Jul 2017 |
Field building and careers |
| Resolved true |
Factual |
21 Jul 2017 |
Dario Amodei's PhD was not in machine learning.
Factual · 21 Jul 2017 |
Their own work or organisation |
| Not researched |
Factual |
21 Jul 2017 |
MIRI does some AI safety work based on mathematics and formal logic.
Factual · 21 Jul 2017 |
AI safety and alignment |
| Not researched |
Factual |
21 Jul 2017 |
Dario Amodei had never worked at a startup before OpenAI.
Factual · 21 Jul 2017 |
Their own work or organisation |
| Unresolvable |
Factual |
21 Jul 2017 |
Statistical AI systems have pretty decisively won out over rule-based systems.
Factual · 21 Jul 2017 |
AI capabilities |
| Resolved true |
Factual |
21 Jul 2017 |
OpenAI is a nonprofit AI research lab.
Factual · 21 Jul 2017 |
Their own work or organisation |
| Not researched |
Factual |
21 Jul 2017 |
OpenAI and DeepMind are probably the AI organisations most focused on reinforcement learning.
Factual · 21 Jul 2017 |
AI capabilities |
| Unresolvable |
Factual |
21 Jul 2017 |
Amodei said he thinks OpenAI has generally been very selective in whom it picks.
Factual · 21 Jul 2017 |
Their own work or organisation |
| Not researched |
Factual |
21 Jul 2017 |
In the method Amodei described, the agent builds a model of what reward function would lie behind the human's expressed preferences.
Factual · 21 Jul 2017 |
AI capabilities |
| Not researched |
Factual |
17 Aug 2015 |
The average psychopath is probably a person with poor economic prospects and poor impulse control who spends significant time in prison, rather than a high-status person as portrayed in media.
Factual · 17 Aug 2015 |
Philosophy, ethics and society |
| Mixed |
Factual |
17 Aug 2015 |
Poorly implemented services are currently a major driver of cynicism about government.
Factual · 17 Aug 2015 |
AI policy and governance |
| Mostly true |
Factual |
17 Aug 2015 |
The impact of nuclear power, supersonic flight and elevators has been substantially reduced by regulations or misplaced fears despite working well technically.
Factual · 17 Aug 2015 |
AI policy and governance |
| Not researched |
Factual |
17 Aug 2015 |
Dario Amodei was an early donor to GiveWell and learned that some health charities are far more effective than others.
Factual · 17 Aug 2015 |
Their own work or organisation |
| Unresolvable |
Factual |
17 Aug 2015 |
Successful biology projects are often scrappy or afterthoughts not initially thought promising, rather than massively funded efforts.
Factual · 17 Aug 2015 |
Science and research methods |
| Not researched |
Factual |
17 Aug 2015 |
Dario Amodei has professional experience in both biology and neuroscience.
Factual · 17 Aug 2015 |
Their own work or organisation |
| Not researched |
Factual |
17 Aug 2015 |
GDP per capita is about $2,000 in Sub-Saharan Africa versus about $75,000 in the United States.
Factual · 17 Aug 2015 |
Economics and markets |
| Not researched |
Factual |
17 Aug 2015 |
The average cancer drug increases survival by a few months while having significant side effects that need careful measurement.
Factual · 17 Aug 2015 |
Science and research methods |
| Not researched |
Factual |
17 Aug 2015 |
The computational democracy project has undertaken early ideas on using AI to aggregate opinions and drive consensus, including collaborations with Anthropic.
Factual · 17 Aug 2015 |
AI policy and governance |
| Mostly true |
Factual |
17 Aug 2015 |
Diseases have been eradicated and many countries have gone from poor to rich, through decisions that exhibit high returns to intelligence.
Factual · 17 Aug 2015 |
Economics and markets |
| Mostly true |
Factual |
17 Aug 2015 |
Although drug development is overly cumbersome, once developed drugs are generally successfully deployed and used.
Factual · 17 Aug 2015 |
Economics and markets |
| Unresolvable |
Factual |
17 Aug 2015 |
In recent years GWAS studies seem to suggest that correlations between psychiatric disease genes and positive traits might have been overstated.
Factual · 17 Aug 2015 |
Science and research methods |
| Unresolvable |
Factual |
17 Aug 2015 |
Over the last 30 years, mathematicians, computer scientists and physicists applying their skills to biology have been quite successful but have not had the truly transformative impact initially hoped for.
Factual · 17 Aug 2015 |
Science and research methods |
| Mixed |
Factual |
17 Aug 2015 |
Human economic planners made the decisions that led to several East Asian economies' late-20th-century growth by pulling a few key levers such as export-led industrial policy, not by directly controlling whole economies.
Factual · 17 Aug 2015 |
Economics and markets |
| Not researched |
Factual |
17 Aug 2015 |
Scientists did not understand how or why neurons fired until the 1950s.
Factual · 17 Aug 2015 |
Science and research methods |
| Unresolvable |
Factual |
17 Aug 2015 |
Markets function better in the developed world and are typically good at bringing down the cost of high-value technologies over time.
Factual · 17 Aug 2015 |
Economics and markets |
| Unresolvable |
Factual |
17 Aug 2015 |
There is perhaps about one major biology discovery per year, and collectively these arguably drive over 50% of progress in biology.
Factual · 17 Aug 2015 |
Science and research methods |
| Mixed |
Factual |
17 Aug 2015 |
Smallpox was fully eliminated in the 1970s, and polio and guinea worm are nearly eradicated with fewer than 100 cases per year.
Factual · 17 Aug 2015 |
Biosecurity |
| Unresolvable |
Factual |
17 Aug 2015 |
Dario Amodei thinks and talks a lot about the risks of powerful AI.
Factual · 17 Aug 2015 |
Their own work or organisation |