| Not researched |
Factual |
28 Feb 2025 |
Dario Amodei has never used the term 'P-Doom'.
Factual · 28 Feb 2025 |
Their own work or organisation |
| Resolved true |
Factual |
28 Feb 2025 |
Anthropic's approach to reasoning models in Sonnet 3.7 emphasized different things from competitors' reasoning models.
Factual · 28 Feb 2025 |
AI capabilities |
| Not researched |
Factual |
28 Feb 2025 |
Reasoning models offered by AI companies have generally been separate from the regular model, as a regular model plus a distinct reasoning model.
Factual · 28 Feb 2025 |
AI capabilities |
| Not researched |
Factual |
28 Feb 2025 |
Amodei said that they found that with "CLAWD" (as transcribed, apparently Claude), the task, described as "a bunch of statistical analysis", can be done in three days.
Factual · 28 Feb 2025 |
AI capabilities |
| Unresolvable |
Factual |
18 Feb 2025 |
Chris Olah was unusual in proposing and seriously pursuing a comprehensive effort to understand everything neural networks do.
Factual · 18 Feb 2025 |
AI safety and alignment |
| Mostly true |
Factual |
18 Feb 2025 |
The lack of understanding of how AI systems work is essentially unprecedented in the history of technology.
Factual · 18 Feb 2025 |
AI safety and alignment |
| Not researched |
Factual |
18 Feb 2025 |
For some applications such as mortgage assessments, where decisions are legally required to be explainable, the inability to see inside AI models is a legal blocker to adoption.
Factual · 18 Feb 2025 |
AI policy and governance |
| Not researched |
Factual |
18 Feb 2025 |
For decades the conventional wisdom was that understanding AI models was impossible and that they were inscrutable black boxes.
Factual · 18 Feb 2025 |
AI safety and alignment |
| Not researched |
Factual |
18 Feb 2025 |
The field of AI as a whole is further ahead than interpretability efforts and is advancing very quickly.
Factual · 18 Feb 2025 |
AI safety and alignment |
| Not researched |
Factual |
18 Feb 2025 |
There is also evidence of real-world behavior that looks vaguely like “cheating on the test”, though it’s more degenerate than it is…
Factual · 18 Feb 2025 |
AI safety and alignment |
| Mixed |
Factual |
18 Feb 2025 |
People outside the AI field are often surprised and alarmed to learn that AI creators do not understand how their AI systems work.
Factual · 18 Feb 2025 |
Philosophy, ethics and society |
| Not researched |
Factual |
18 Feb 2025 |
there are a lot of experiments (many of which were done by Anthropic) showing that models can lie or deceive under certain circumstances…
Factual · 18 Feb 2025 |
AI safety and alignment |
| Not researched |
Factual |
18 Feb 2025 |
Anthropic's interpretability team immediately found basic mechanisms in a language model for copying and sequential pattern-matching.
Factual · 18 Feb 2025 |
AI safety and alignment |
| Not researched |
Factual |
18 Feb 2025 |
What there isn’t is evidence of dangerous behaviors emerging in a more naturalistic way, or of a general tendency or general intent to lie…
Factual · 18 Feb 2025 |
AI safety and alignment |
| Mostly true |
Factual |
18 Feb 2025 |
When Anthropic was founded, its founders made mechanistic interpretability a central part of the company's direction, focused on LLMs.
Factual · 18 Feb 2025 |
Their own work or organisation |
| Mostly true |
Factual |
18 Feb 2025 |
Amodei said that "we" were able to provide a strong theoretical basis for having conviction that superposition was the right hypothesis to pursue; the surrounding text does not say who "we" refers to.
Factual · 18 Feb 2025 |
AI safety and alignment |
| Mostly true |
Factual |
18 Feb 2025 |
It became clear that superposition would be a central problem for language models, much worse than in vision models.
Factual · 18 Feb 2025 |
AI safety and alignment |
| Not researched |
Factual |
18 Feb 2025 |
we’ve never seen any solid evidence in truly real-world scenarios of deception and power-seeking
Factual · 18 Feb 2025 |
AI safety and alignment |
| Mostly true |
Factual |
18 Feb 2025 |
Modern generative AI systems are opaque in a way that fundamentally differs from traditional software.
Factual · 18 Feb 2025 |
AI safety and alignment |
| Unresolvable |
Factual |
18 Feb 2025 |
Amodei said he does not think truly compelling evidence of the dangers of autonomous AI exists today, contrary to the claims of advocates.
Factual · 18 Feb 2025 |
AI safety and alignment |
| Not researched |
Factual |
18 Feb 2025 |
One year before this essay, neural network thoughts could not be traced and millions of concepts inside them could not be identified; as of the essay, they can.
Factual · 18 Feb 2025 |
AI safety and alignment |
| Resolved true |
Factual |
18 Feb 2025 |
Anthropic created "Golden Gate Claude", a version of one of its models with the Golden Gate Bridge feature artificially amplified, making it obsessed with the bridge.
Factual · 18 Feb 2025 |
AI safety and alignment |
| Not researched |
Factual |
18 Feb 2025 |
A very popular mechanistic interpretability workshop at the ICML conference was reportedly rejected on seemingly pretextual grounds.
Factual · 18 Feb 2025 |
Field building and careers |
| Not researched |
Factual |
18 Feb 2025 |
When a generative AI system performs a task, researchers do not know at a specific or precise level why it makes the choices it does.
Factual · 18 Feb 2025 |
AI safety and alignment |
| Not researched |
Factual |
18 Feb 2025 |
Anthropic found over 30 million features in Claude 3 Sonnet, a medium-sized commercial model.
Factual · 18 Feb 2025 |
AI safety and alignment |
| Mostly true |
Factual |
18 Feb 2025 |
Circuits, groups of features, allow researchers to trace an AI model's thinking steps.
Factual · 18 Feb 2025 |
AI safety and alignment |
| Not researched |
Factual |
18 Feb 2025 |
AI models use superposition because it lets them express more concepts than they have neurons, enabling them to learn more.
Factual · 18 Feb 2025 |
AI safety and alignment |
| Not researched |
Factual |
18 Feb 2025 |
Even a small AI model may contain a billion or more concepts.
Factual · 18 Feb 2025 |
AI capabilities |
| Not researched |
Factual |
18 Feb 2025 |
Without interpretability, the only way to discover that a jailbreak of an AI model exists is to find it empirically.
Factual · 18 Feb 2025 |
AI safety and alignment |
| Not researched |
Factual |
18 Feb 2025 |
Research papers from the preceding few months have shown that interpretability can help humans understand patterns predicted by AI in DNA and protein sequence data.
Factual · 18 Feb 2025 |
Science and research methods |
| Resolved true |
Factual |
11 Nov 2024 |
Anthropic's Responsible Scaling Policy uses an if-then structure: if models pass a certain capability, Anthropic imposes a set of safety and security requirements.
Factual · 11 Nov 2024 |
AI safety and alignment |
| Not researched |
Factual |
11 Nov 2024 |
Dario Amodei said he believes Claude Haiku 3.5, the smallest new model, is about as good as Opus 3, the largest old model.
Factual · 11 Nov 2024 |
AI capabilities |
| Own records only |
Factual |
11 Nov 2024 |
Around 2016 or 2017, Dario Amodei first came to believe in, or confirmed his belief in, the scaling hypothesis.
Factual · 11 Nov 2024 |
AI capabilities |
| Own records only |
Factual |
11 Nov 2024 |
With not much additional training, Claude models became quite good at computer use tasks.
Factual · 11 Nov 2024 |
AI capabilities |
| Own records only |
Factual |
11 Nov 2024 |
Models becoming lazy in coding is not due to Anthropic trying to save on compute or to seasonal effects, but to the difficulty of steering model behaviour.
Factual · 11 Nov 2024 |
Their own work or organisation |
| Not researched |
Factual |
11 Nov 2024 |
Anthropic's model training includes a post-training phase using reinforcement learning from human feedback and other kinds of reinforcement learning.
Factual · 11 Nov 2024 |
AI capabilities |
| Not researched |
Factual |
11 Nov 2024 |
Dario Amodei said that he and Ilya Sutskever were "probably" the ones who "really kind of set the research direction" at the organisation where he was vice president of research.
Factual · 11 Nov 2024 |
Their own work or organisation |
| Not researched |
Factual |
11 Nov 2024 |
In the latest RSP version, released in the prior month or two, Anthropic tests autonomy risk by the model's ability to do aspects of AI research itself.
Factual · 11 Nov 2024 |
AI safety and alignment |
| Not researched |
Factual |
11 Nov 2024 |
The current Claude computer-use model makes mistakes and misclicks.
Factual · 11 Nov 2024 |
AI capabilities |
| Own records only |
Factual |
11 Nov 2024 |
Anthropic sandboxes its models during training, for example not exposing the model to the internet.
Factual · 11 Nov 2024 |
AI safety and alignment |
| Mostly true |
Factual |
11 Nov 2024 |
Anthropic released computer use first in API form rather than to consumers, partly for safety reasons.
Factual · 11 Nov 2024 |
Their own work or organisation |
| Not researched |
Factual |
11 Nov 2024 |
Amodei has worked in the AI field for about 10 years.
Factual · 11 Nov 2024 |
Their own work or organisation |
| Own records only |
Factual |
11 Nov 2024 |
Amodei has seen no evidence so far that a new optimization method or technique is needed to unblock AI progress.
Factual · 11 Nov 2024 |
AI timelines and forecasting |
| Mostly true |
Factual |
11 Nov 2024 |
Anthropic's Responsible Scaling Policy does not specify ASL-4 requirements until ASL-3 has been reached.
Factual · 11 Nov 2024 |
AI safety and alignment |
| Not researched |
Factual |
11 Nov 2024 |
As of November 2024, Anthropic's current models are classified ASL-2 under its Responsible Scaling Policy.
Factual · 11 Nov 2024 |
Their own work or organisation |
| Own records only |
Factual |
11 Nov 2024 |
Anthropic today makes a model and then deploys thousands, maybe tens of thousands, of instances of it.
Factual · 11 Nov 2024 |
AI compute, hardware and data |
| Not researched |
Factual |
11 Nov 2024 |
Anthropic's Haiku is its small, fast, cheap model, which at release was surprisingly intelligent for its speed and cost.
Factual · 11 Nov 2024 |
AI capabilities |
| Not researched |
Factual |
11 Nov 2024 |
Amodei and colleagues have documented scaling laws in many domains other than language.
Factual · 11 Nov 2024 |
AI capabilities |
| Unresolvable |
Factual |
11 Nov 2024 |
As of late 2024, pre-training accounts for the majority of the cost of building AI models.
Factual · 11 Nov 2024 |
AI compute, hardware and data |
| Not researched |
Factual |
11 Nov 2024 |
Dario Amodei was vice president of research at OpenAI for roughly the last couple of years of his roughly five years there.
Factual · 11 Nov 2024 |
Their own work or organisation |
| Own records only |
Factual |
11 Nov 2024 |
Amodei observed in his early AI work that models performed better with more data, larger size and longer training.
Factual · 11 Nov 2024 |
AI capabilities |
| Not researched |
Factual |
11 Nov 2024 |
Performance on a programming benchmark has gone from succeeding about 3% of the time to about 50% of the time.
Factual · 11 Nov 2024 |
AI capabilities |
| Not researched |
Factual |
11 Nov 2024 |
Anthropic is not currently building its own coding IDEs but powers companies such as Cursor and Kognition that build them on its API.
Factual · 11 Nov 2024 |
Their own work or organisation |
| Not researched |
Factual |
11 Nov 2024 |
Anthropic's interpretability research had no commercial application for three or four years and still has none.
Factual · 11 Nov 2024 |
Their own work or organisation |
| Resolved true |
Factual |
11 Nov 2024 |
On some computer-use benchmarks, previous models succeeded about 6% of the time and Anthropic's new model succeeds 14 or 22% of the time.
Factual · 11 Nov 2024 |
AI capabilities |
| Not researched |
Factual |
11 Nov 2024 |
OpenAI built GPT-3, the model that was commercialized, and Dario Amodei was involved in its commercialization.
Factual · 11 Nov 2024 |
Their own work or organisation |
| Not researched |
Factual |
11 Nov 2024 |
Anthropic was the only AI company that felt positively about California's SB 1047 in a very detailed way, while some of the big companies, such as Google, OpenAI, Meta and Microsoft, were pretty staunchly against it.
Factual · 11 Nov 2024 |
AI policy and governance |
| Not researched |
Factual |
11 Nov 2024 |
Asked about the limits and expense of compute, Amodei guessed that most frontier model companies are operating at roughly a 1 billion scale, plus or minus a factor of three.
Factual · 11 Nov 2024 |
Economics and markets |
| Not researched |
Factual |
11 Nov 2024 |
Claude's computer use feature adds that input images can be computer screenshots, with the model trained to output screen click locations or key presses.
Factual · 11 Nov 2024 |
AI capabilities |
| Resolved true |
Factual |
11 Nov 2024 |
Anthropic's updated Claude Sonnet 3.5 scores about 50% on SWE-bench.
Factual · 11 Nov 2024 |
AI capabilities |
| Not researched |
Factual |
11 Nov 2024 |
Sonnet 3.5 has roughly the same cost and speed as Sonnet 3 but is smarter than the original Opus 3, especially for code.
Factual · 11 Nov 2024 |
AI capabilities |
| Not researched |
Factual |
11 Nov 2024 |
Anthropic has had to change its Responsible Scaling Policy framework, releasing a new version a few weeks before November 2024.
Factual · 11 Nov 2024 |
Their own work or organisation |
| Mostly true |
Factual |
11 Nov 2024 |
Anthropic wants to evaluate every new model to see whether it is getting close to dangerous (CBRN) capabilities, and does not think current models yet pose these risks seriously.
Factual · 11 Nov 2024 |
Biosecurity |
| Mostly true |
Factual |
11 Nov 2024 |
Amodei first joined the AI field in late 2014, working at Baidu with Andrew Ng.
Factual · 11 Nov 2024 |
Their own work or organisation |
| Not researched |
Factual |
11 Nov 2024 |
With each new model, every few months, Anthropic measures that its models are getting better at concerning tasks as well as at valuable, economically useful tasks.
Factual · 11 Nov 2024 |
AI safety and alignment |
| Resolved true |
Factual |
11 Nov 2024 |
Anthropic has almost a thousand employees, and its risk policy is meant to make sure they understand that taking these risks seriously is one of the company's highest priorities, if not the highest.
Factual · 11 Nov 2024 |
Their own work or organisation |
| Unresolvable |
Factual |
11 Nov 2024 |
There are probably about 100 people working on the research area being discussed, but not around 10,000.
Factual · 11 Nov 2024 |
AI safety and alignment |
| Own records only |
Factual |
11 Nov 2024 |
For the most part, with some narrow exceptions, deployed models are not changing after release.
Factual · 11 Nov 2024 |
Their own work or organisation |
| Not researched |
Factual |
11 Nov 2024 |
Productivity increases from the computer and internet revolutions were generally underwhelming, less than might be imagined.
Factual · 11 Nov 2024 |
Economics and markets |
| Own records only |
Factual |
11 Nov 2024 |
Anthropic has never changed the weights of a deployed model without telling anyone.
Factual · 11 Nov 2024 |
Their own work or organisation |
| Not researched |
Factual |
11 Nov 2024 |
Anthropic showed demos of Claude filling out spreadsheets, interacting with websites and opening programs on Windows, Linux and Mac.
Factual · 11 Nov 2024 |
AI capabilities |
| Own records only |
Factual |
11 Nov 2024 |
In 2017, seeing GPT-1 results convinced Amodei that language was probably the area where scaling could work.
Factual · 11 Nov 2024 |
Their own work or organisation |
| Not researched |
Factual |
11 Nov 2024 |
On typical real-world programming tasks, AI models went from 3% in January 2024 to 50% in October 2024.
Factual · 11 Nov 2024 |
AI capabilities |
| Resolved true |
Factual |
11 Nov 2024 |
DeepMind's AlphaGo Zero required no example data from humans.
Factual · 11 Nov 2024 |
AI capabilities |
| Mostly true |
Factual |
12 Apr 2024 |
The Guns of August describes how the crisis erupted quickly and almost no one grasped the full picture.
Factual · 12 Apr 2024 |
Other |
| Own records only |
Factual |
12 Apr 2024 |
Amodei says he has found a workflow that does not use the models for the part where they help little (in the surrounding text, generating initial ideas), and says they can help with expressing and refining ideas.
Factual · 12 Apr 2024 |
Their own work or organisation |
| Own records only |
Factual |
12 Apr 2024 |
Dario Amodei occasionally uses Anthropic's internal models to assist with his strategy writing and thinking.
Factual · 12 Apr 2024 |
Their own work or organisation |
| Mostly true |
Factual |
12 Apr 2024 |
Anthropic's experiment found its large model nearly matched a hired human team in changing people's views with persuasive essays.
Factual · 12 Apr 2024 |
AI safety and alignment |
| Own records only |
Factual |
12 Apr 2024 |
In Amodei's experience, the company's internal models he occasionally uses can in some respects help conceptualize thinking but are not good at generating initial ideas.
Factual · 12 Apr 2024 |
AI capabilities |
| Unresolvable |
Factual |
12 Apr 2024 |
Amodei said Anthropic launched its safety-level framework in September and that about three months later OpenAI launched a similar product.
Factual · 12 Apr 2024 |
AI policy and governance |
| Not researched |
Factual |
12 Apr 2024 |
The Guns of August is a history book about the causes of the outbreak of World War I.
Factual · 12 Apr 2024 |
Other |
| Resolved true |
Factual |
12 Apr 2024 |
The Star Trek series depicts a highly developed universe in which humans live longer and have expanded into space.
Factual · 12 Apr 2024 |
Other |
| Not researched |
Factual |
12 Apr 2024 |
Current AI models cost about $100 million to train, give or take two to three times.
Factual · 12 Apr 2024 |
AI compute, hardware and data |
| Own records only |
Factual |
12 Apr 2024 |
Dario Amodei said that at Anthropic he writes strategic planning documents or does some new thinking, possibly on directions others have not covered.
Factual · 12 Apr 2024 |
Their own work or organisation |
| Mostly true |
Factual |
12 Apr 2024 |
Anthropic introduced its Responsible Scaling Plan at the end of 2023, setting thresholds for AI system capabilities.
Factual · 12 Apr 2024 |
Their own work or organisation |
| Not researched |
Factual |
12 Apr 2024 |
Dario Amodei worked at OpenAI for five years and was one of its first employees.
Factual · 12 Apr 2024 |
Their own work or organisation |
| Mixed |
Factual |
12 Apr 2024 |
我们和其他一些研究者发现,模型里有时会有某些特定的神经元和统计指标(这些通常不出现在它的外部反应中),这些指标能够指示模型是在撒谎还是在说真话。
Factual · 12 Apr 2024 |
AI safety and alignment |
| Not researched |
Factual |
12 Apr 2024 |
The Star Trek universe still features geopolitical problems, inequality and exploitation, sometimes worse than today.
Factual · 12 Apr 2024 |
Other |
| Not researched |
Factual |
12 Apr 2024 |
In 2018 OpenAI developed GPT-1, which used about one hundred-thousandth of the compute of current models.
Factual · 12 Apr 2024 |
AI compute, hardware and data |
| Unresolvable |
Factual |
12 Apr 2024 |
Amodei and colleagues found regular scaling patterns and predicted that raising investment from about $10,000 to hundreds of millions of dollars would bring huge change and economic value.
Factual · 12 Apr 2024 |
AI capabilities |
| Unresolvable |
Factual |
12 Apr 2024 |
Anthropic's Claude 3 Opus reaches roughly 99.9% accuracy on 20-digit addition, well beyond what common internet data would give.
Factual · 12 Apr 2024 |
AI capabilities |
| Not researched |
Factual |
12 Apr 2024 |
Kennedy reportedly read The Guns of August before the Cuban Missile Crisis.
Factual · 12 Apr 2024 |
Other |
| Resolved true |
Factual |
12 Apr 2024 |
Anthropic's models are currently at AI Safety Level ASL-2.
Factual · 12 Apr 2024 |
Their own work or organisation |
| Own records only |
Factual |
12 Apr 2024 |
Anthropic has already secured the compute it needs for 2024, and Amodei expects next year's compute to be sufficient too.
Factual · 12 Apr 2024 |
AI compute, hardware and data |
| Own records only |
Factual |
12 Apr 2024 |
In Amodei's short-term personal experience, he has always found a workflow in which AI tools assist him.
Factual · 12 Apr 2024 |
Their own work or organisation |
| Not researched |
Factual |
8 Aug 2023 |
When a model is trained to be aligned, researchers do not know what happens inside the model.
Factual · 8 Aug 2023 |
AI safety and alignment |
| Unresolvable |
Factual |
8 Aug 2023 |
Dario Amodei did not think AI scaling would hit a wall, and a few other people shared that view, when a wall was a more plausible case.
Factual · 8 Aug 2023 |
Their own work or organisation |
| Not researched |
Factual |
8 Aug 2023 |
Amodei said he thought multiple companies were already in the range of $100 million to $1 billion per year.
Factual · 8 Aug 2023 |
Economics and markets |
| Unresolvable |
Factual |
8 Aug 2023 |
As of 2023, the model can be very good at coding but can't quite prove the prime number theorem yet.
Factual · 8 Aug 2023 |
AI capabilities |
| Own records only |
Factual |
8 Aug 2023 |
Anthropic has generally found that hired physics PhDs can learn machine learning and contribute very quickly in most cases.
Factual · 8 Aug 2023 |
Field building and careers |