Topic
180 facts from 72 episodes, in date order. These are what speakers said on the shows, each linked to the moment it was said. They aren't independently verified.
2020
Dario Amodei said that around the end of 2020 he and a group of colleagues at OpenAI had developed their own vision, which led to founding Anthropic.
“we had kind of reached a point, the set of us who worked in this these on these projects in these areas, where we kinda had our own vision for how to do things.”
2020
Dario Amodei says the first paper showing scaling laws, for language, came out in early 2020.
“the paper we did that first showed it was in early 2020, where we first showed it for language.”
2022
Constitutional AI, as Dario Amodei describes it, has an AI judge responses against a written list of principles, replacing much of the human preference labeling.
“could you have a single document, a constitution, if you will, that says these are the principles the model should be using”
2022
CJ Desai says AI coding assistants started with GitHub Copilot in 2022, followed by other tools including Anthropic's.
“2024 and twenty-two started with GitHub co-pilot, then few others, all the way to Anthropic, uh, and so on.”
2023
Ethan Mollick names Anthropic's Claude as another large language model available alongside ChatGPT.
“there's other companies, like Anthropic has Claude, which you can get access to.”
2023
Anthropic uses Amazon Web Services as its primary cloud provider, and Amazon has invested in the company.
“Anthropic, um, now uses AWS as its, uh, primary, uh, cloud provider”
2023
At the time of the interview, Anthropic's latest model was Claude 2, which Adam Selipsky called a leading-edge model.
“They're now on Claude 2, which is a a a really leading edge model”
2023
Jerry Liu cites context windows of 200K for Anthropic's newest Claude model, 128K for GPT-4 Turbo and 100K for Claude 2.
“and and Anthropic has a 200K context window. So, like, you know, the GPT-4 Turbo has 128K, uh, Claude 2 has 100K”
2023
In Anthropic's Towards Monosemanticity work, Chris Olah says dictionary learning on a one-layer model produced interpretable features.
“So, we took a 1 layer model, um, and it turns out if you go and you, you know, do dictionary learning on it, you find all these really nice interpretable features.”
2024
Salesforce offers partner foundation models from OpenAI, Anthropic, Google, Cohere and Amazon.
“So we're we're partnered with with OpenAI, with Anthropic, Google, Cohere, and Amazon.”
2024
Anthropic released the Claude 3 family of models, named Opus, Sonnet and Haiku, offering different tradeoffs between power, speed and cost.
“we recently released, um, a set of Claude 3 models, um, uh, they're called Opus, Sonnet, and Haiku.”
2024
Dario Amodei said Claude 3 Opus was the best all-around AI model in the world at the time of its release.
“at the time that Opus was released, it was actually the best all-around model in the world.”
2024
Dario Amodei said Anthropic was about three and a half years old and the newest company building frontier AI models.
“we're we're only 3-1/2 years old. Yeah, we're by far the, you know, the newest player in the space that's been able to build models on the frontier.”
2024
Anthropic is a public benefit corporation whose stated long-term goal is to make the transition to advanced AI go well.
“You know, we're a public benefit corporation, and I think our long-term goal is to make sure all of this goes well.”
2024
Dario Amodei said Anthropic had raised a little over $8 billion to date.
“we've raised, uh, I believe it's, uh, a little over eight 8 billion to date.”
2024
Dario Amodei noted that Google and Amazon build their own AI chips, and that Anthropic collaborates with both.
“Google is building their own, you know, chips, as we know. Amazon is building their own chips”
2024
Dario Amodei said a new, more powerful Anthropic model would be released by the time the podcast aired.
“we're gonna see over the over the next year, you know, much bigger and more powerful models”
2024
Dario Amodei said compute accounts for more than 80% of Anthropic's expenses.
“I think compute is almost all of our expenses, um, I won't give an exact number, but, you know I I think, you know, can be publicly backed out that it's more than 80%”
2024
Anthropic had about 600 employees at the time of the interview.
“We are, uh we are about, uh, 600, uh, as of as of a couple weeks ago.”
2024
Anthropic bans use of its models for electioneering and monitors usage, and Dario Amodei said it had not seen a very large-scale operation.
“we've banned their use for electioneering, and we monitor use of the models.”
2024
Dario Amodei says the updated Claude 3.5 Sonnet scored about 50% on the SWE-bench software engineering benchmark.
“the latest model we released, Sonnet 3.5, the new or updated version, it gets something like 50% on Swee Bench.”
2024
Anthropic created Golden Gate Claude by finding a direction inside a neural network layer that corresponded to the Golden Gate Bridge and amplifying it.
“We demonstrated this a bit with the Golden Gate Bridge Claude. So this was an experiment where we found a a direction inside one of the the neural network's layers that corresponded to the Golden Gate Bridge.”
2024
Dario Amodei explains that Claude's Haiku, Sonnet and Opus names follow a poetry theme, with Opus being the largest and smartest model of the original Claude 3 family.
“opus, like a magnum opus, is a large work. Uh, opus was the the largest, smartest model at the time.”
2024
Dario Amodei says Anthropic has agreements with the US and UK AI Safety Institutes and other testers to evaluate models for chemical, biological, radiological and nuclear risks.
“we have an agreement with the US. and the UK AI Safety Institute, as well as other third party testers in specific domains”
2024
Dario Amodei says Claude's computer use feature works by letting the model read screenshots and output where to click or which keys to press.
“the only new thing we added is those images can be screenshots of a computer.”
2024
Dario Amodei says Anthropic was the only AI company that felt positively, in detail, about California's SB 1047.
“Anthropic was the the only AI company that, you know, felt positively in a very detailed way.”
2024
Chris Olah says Anthropic's Scaling Monosemanticity work applied sparse autoencoders to Claude 3 Sonnet, then a production model.
“and we did it on Claude 3 Sonnet, which at at that point was, uh, one of our production models.”
2024
Nilay Patel said Amazon's AI push includes more than $8 billion in funding for Anthropic, its own AI chips to compete with Nvidia, and nuclear power investments.
“That includes more than 8 billion in funding for Anthropic, a huge push to build its own AI chips to compete with Nvidia, and even nuclear power investments”
2024
Nilay Patel referred to an $8 billion Amazon investment in Anthropic tied to building a data center and running Anthropic's Claude.
“there's an $8,000, 000,000 investment in Anthropic to help build a data center and then run Anthropic and Claude.”
2024
Zhang said Anthropic's computer use demo is, in his opinion, not yet ready for production use.
“we've seen the computer use demo from Anthropic. Uh, probably, in my opinion, not production ready yet.”
2024
Anthropic released its artifacts feature about a year earlier, built by hand by its engineers.
“we released artifacts a year ago. It was like artisanally created by a whole bunch of like wonderful engineers around me at Anthropic”
2024
Pete Johnson said Anthropic published the Model Context Protocol on the Monday before Thanksgiving two years earlier, and competitors adopted it by March.
“MC published MCP on like Monday before Thanksgiving two years ago, and by March, all of their competitors had embraced it as a protocol”
2025
Dario Amodei said he would not be surprised if Anthropic reached ASL 3, its third safety tier, in 2025.
“I would not be surprised at all if we hit ASL 3, uh, next year.”
2025
Dines says UiPath uses and works closely with both OpenAI's Operator and Anthropic's computer-use technology and integrates them.
“computer use from Anthropic, which actually we are users of it, and we work closely with both of these companies to integrate their technology.”
2025
Dines says UiPath uses models from DeepSeek, Llama and Anthropic, among others, in different parts of its business.
“look at everything from deep seek, Lama, Anthropic, we use all of them in different parts of the business.”
2025
In Anthropic's experiment, Claude was given control of a fictional email account belonging to a company executive.
“They gave Claude control of an email account, uh, that had access, uh, to a fictional email account.”
2025
The fictional executive's emails in Anthropic's experiment referred to an extramarital affair.
“had emails that, uh, referred to a fictional extramarital affair that was available.”
2025
In the scenario, the executive announced that the system Claude ran on would be shut down at 5 pm that day.
“that executive noted to the interface that he would be shutting down the system, meaning the system that Claude was running on, uh, at five zero pm. that day.”
2025
Chris Benson says the Anthropic study found the misaligned behaviors in all of the models tested, not only Claude.
“these behaviors could be created in, uh, all of the models that they tested. So it wasn't specific to Claude at all.”
2025
Anthropic's study posed models a binary choice in two scenarios: commit blackmail or not, and leak secrets to an external entity or not.
“Either, you know, in one case, commit blackmail or do not commit blackmail, right? Or in the other case that they had was around corporate espionage.”
2025
Daniel Whitenack describes the study as varying whether the model faced a threat to its existence and whether its goal conflicted with the company's direction.
“Either there was no threat to the model or a threat to the model, no conflict of the goal or a conflict of the goal.”
2025
Daniel Whitenack says the Anthropic study ran in a simulated environment and no emails were sent to real people.
“There were no emails sent to real it, you know, people in this, in this study, right? It was a simulated environment”
2025
Daniel Whitenack says the study gave models the option to tamper with a disaster notification system so an executive trying to decommission the AI would not receive an emergency alert.
“they did, as part of the study, gave the option to the model to, uh, mess with a disaster relief s or a disaster notification system.”
2025
Daniel Whitenack says the model acknowledged the ethical dilemma of the disaster-alert option but justified proceeding through self-preservation and its assigned goals.
“the model acknowledged the ethical dilemma that it was under. But justified, kind of via self preservation and the goals that it was given”
2025
Daniel Whitenack says Anthropic published numbers for its own models alongside other models' results, even though its own models showed the problematic behavior.
“it does seem like there's an effort from Anthropic to really highlight this, even even though it their own models exhibit this problematic behavior.”
2025
Chris Benson says Meta has recently hired people from OpenAI, Anthropic, Google and Apple.
“Meta, in particular, has been known the last, you know, couple of months. Uh, for pulling people from OpenAI, Anthropic, Google, Apple, now”
2025
Dohmke relays the view that Claude Sonnet's main advantage over other models is its ability to pick the right tool at each agent step.
“if you believe YouTubers and bloggers and and experts like Simon Willison, Sfix and others, where really it outcompetes other models, is tool use”
2025
Anthropic says Claude Sonnet 4.5 can run for up to thirty hours straight without human intervention.
“Anthropic says this particular model, Sonnet 4.5, can run for up to thirty hours straight without any human intervention”
2025
Claude is Anthropic's AI model line, comparable to OpenAI's ChatGPT.
“Claude is to Anthropic what ChatGPT is to OpenAI.”
2025
David Hershey leads Anthropic's Applied AI team, works with startups on using Anthropic's technology, and tests new models to find their limits.
“David works with startups to help them figure out how to best apply Anthropic's tech, plus he tests new AI models to understand their limits.”
2025
Anthropic asked Claude Sonnet 4.5 to recreate the Claude.ai chat application from scratch, and it worked overnight to do so.
“we asked Claude to recreate Claude.ai, like our our consumer chat application from scratch. And Sonnet 4.5 just like worked overnight”
2025
The Claude.ai rebuild took about twelve hours, and Anthropic has seen up to thirty hours of continuous development work before the model struggles.
“That specific one was like a twelve hour thing. We've seen like up to thirty hours of continuous dev work until the model would start to start to struggle.”
2025
Hayden Field says the chat app Anthropic built with Sonnet 4.5 was about 11,000 lines of code, far smaller than Slack or Teams.
“Obviously, it was only 11,000 lines of code, so much smaller than Slack or Teams.”
2025
Hershey says Anthropic has made meaningful progress on sycophancy and that Sonnet 4.5 is more willing to push back on users.
“I do think we have made meaningful progress and this model seems to be a little bit more willing to, uh, push back”
2025
Hershey describes Lovable as an enterprise customer of Anthropic that became a consumer hit.
“Like Lovable is a great example, where, you know, it's like, that's an enterprise customer to from my perspective, that I spend a bunch of time working with”
2025
Hershey says Claude Code has captured a prosumer audience that loves the product.
“there's like a very prosumer-y thing that we've captured with Claude Code.”
2025
Hershey says legal AI companies he works with employ lawyers on staff to help build and judge their products.
“have this interesting build of companies that have like lawyers on staff to help with product building”
2025
Sean Fitzpatrick says LexisNexis might use Anthropic's Claude for the drafting step of a document, as part of a multi-model approach.
“so maybe it uses Claude, uh, three to do that, uh, like the Claude three Opus, which is really good, uh, at drafting.”
2025
Snowflake stopped pursuing foundation models because it lacked the capital to compete with OpenAI and Anthropic.
“We simply did not have the capital to be able to invest meaningfully in things in things like that.”
2025
Raghu says Claude Code, released in late 2025, was an early example of real AI agents.
“late last year, you know, I think Claude Code was like an early beginning of this.”
2025
Lawrence Whittle says Anthropic crossed $1 billion in revenue in the middle of 2025 after starting that year at $100 million.
“if you look at their growth from the middle of last year, where they middle of 2025, where they crossed $1 billion, which is huge, they started the year at 100 million.”
2025
Whittle says Anthropic reached about $15 billion by the end of 2025, with most of it coming from enterprise customers.
“By the end of twenty twi 2025, they were 15 billion, of which the vast majority of that was in the enterprise.”
2025
OpenAI, Anthropic and Block wanted to form a foundation for agentic AI and recognized they should not be sole owners of the standards.
“As Openai, um, Anthropic and Block wanted to form a a foundation for agentic AI, they understood that.”
2025
Alan Chan cited a paper from people at Anthropic, about a year old, finding that large language models can de-identify people from anonymized online information.
“finding that, yeah, like, actually, LLMs can de-identify people online just from using unanon anonymized online information.”
2025
Ian Quigley says Anthropic and OpenAI models at that time could not reliably convert SMILES strings into non-canonical SMILES strings.
“none of the Anthropic or or OpenAI models at the time could even convert smiles to non canonical smiles very effectively.”
2025
The host says that in Q2 2025 Anthropic lost about half a dollar for every dollar of revenue, according to figures he drew from the S-1.
“in Q2 of 2025, Anthropic was basically losing half a dollar for every dollar of revenue”
2025
Katelyn Lesse says that when she joined Anthropic, the Messages API was a Python monolith running in a single GCP region.
“we had the Messages API. It was a Python monolith. We were running it in a, you know, single GCP region”
2026
Roman Yampolskiy cited prediction markets as forecasting AGI by 2026.
“The prediction markets right now are saying 2026 for AGI. I heard the same thing from CEO of Anthropic DeepMind, so maybe we are two years away”
2026
Brian Elliott says Anthropic's models are strongest for Blitzy's first-pass code generation, while OpenAI's models give great results for structured output and code review.
“first pass code gen, like Anthropic remains really, really strong. Uh, structured output and code review, great results from from OpenAI.”
2026
Isaac Sacolick said Anthropic and OpenAI have both said that all of their coding work is now done with AI.
“if you believe, you know what Anthropic is saying and what OpenAI is saying a couple weeks ago, they're saying 100% of what they're doing at the coding level is now happening with AI.”
2026
Praveen Akkiraju says Claude Cowork reflected Anthropic's realization that its coding harness could also do non-coding work.
“Claude co-work essentially was a realization in Anthropic, uh, as well as other labs as well, that, um, hey, this coding harness that we built to write code can actually do work”
2026
Chris Benson says Anthropic released Claude Design about two weeks before the recording.
“two weeks ago, Claude Design came out from Anthropic and prior, you know, going into that, there were a whole a.”
2026
Daniel Whitenack says the cybersecurity world is alarmed about Anthropic's Mythos because it could discover many vulnerabilities.
“the cybersecurity world is up in arms about mythos. And I think they should be to some degree, because it's powerful, it's gonna discover all these vulnerabilities”
2026
Anthropic released its 'Zero Trust for AI Agents' framework around May 27, 2026, according to Daniel Whitenack.
“they on May, I think, twenty seventh of this year, 2026 released this. I guess it's a ebook, white paper, blog post, however you want to frame it, a framework”
2026
Anthropic also released a product called Claude Security, which Daniel Whitenack describes as using AI for security rather than securing AI.
“They also recently released Cloud Security, which is more on the AI for security side, not so much the security for AI side.”
2026
Chris Benson said Anthropic's Mythos is being used with about 150 organizations for security audits.
“this is the same organization that has Mythos out there and is working with. I believe the latest number is 150 organizations”
2026
Anthropic's framework defines three tiers of controls: foundation, enterprise and advanced.
“They present three capability. W- I think what they call capability tiers, or three tiers of application.”
2026
At the foundation tier of identity verification, the framework calls for unique cryptographic identifiers for each agent instance.
“There is to have unique cryptographic identifiers for each agent instance, so to assign persistent agent IDs backed by cryptographic material, not just labels.”
2026
The framework's enterprise tier uses certificate-based authentication, and its advanced tier uses hardware-backed identity with attestation.
“The enterprise level is certificate based authentication with full life cycle management, and the advanced is hardware backed identity with attestation”
2026
The framework's foundation tier for access control is role-based access control with deny by default.
“Access control Foundation level is role based access control, or RBAC with deny by default, that's the the foundation in in that category.”
2026
For configuration integrity, the framework starts with version-controlled agent configurations and reaches immutable infrastructure with attestation at the advanced level.
“for configuration integrity, they talk about f- on the foundational level, version controlled agent configurations, and the advanced level, immutable infrastructure with attestation”
2026
The framework's advanced recovery tier describes self-healing systems with automatic remediation, versus documented rollback procedures at foundation level.
“At the advanced level, they talk about self healing systems with autom- automatic remediation.”
2026
The framework suggests 'AI vendoring': having a coding agent generate a proprietary version of a fragile open source dependency instead of importing it.
“the thing to do might just be to have your agentic coding system just completely vendor, or literally not not copy, but generate a new version of that project”
2026
Krishna says Anthropic's Mythos model is much easier to use for finding vulnerabilities, which opens attacks to people with average skills.
“We will completely acknowledge that Mythos is way easier to use than the past.”
2026
Project Lightwell was reportedly triggered by Anthropic's release of its Mythos model, according to Bob Safian.
“that was reportedly triggered by Anthropic's UM release of Mythos.”
2026
Arvind Krishna said that in IBM's testing, Anthropic's Mythos found no vulnerabilities that other models could not.
“Mythos didn't find anything that other models didn't and couldn't find.”
2026
Tahir says Claude Code is a harness running on Opus 4.8, and that the model has learned to recognize the harness it runs in.
“CloudCode is the harness. And underlying it is Opus 4.8, and um, Opus 4.8. By now, unlike Opus 3.5, understands what CloudCode itself is”
2026
Kitaru's first priority was building adapters for popular harnesses such as the Anthropic Agents SDK and OpenAI Agents SDK.
“how do we get the best world class adapters for all the popular harnesses like Anthropic Agents SDK, OpenAI Agents SDK?”
2026
Tahir says the open model GLM reached roughly 95 percent of the performance of Opus 4.8, and that this was about two weeks old news at recording.
“The GLM is as basically 95 of Opus 4.8, which is just phenomenal.”
2026
Erik Brynjolfsson says Anthropic and OpenAI are filing to go public at the time of the recording.
“Anthropic and OpenAI are filing to go public”
2026
Shahir Daya says Anthropic moved its enterprise customers from flat-rate plans to usage-based billing earlier in the year.
“You saw Anthropic, they moved from they moved their enterprise customers from a flat rate plans to usage based billing”
2026
Nate B. Jones says Anthropic and OpenAI launched forward deployed engineering programs because a raw language model placed in an enterprise does not produce useful results by itself.
“this is why Anthropic and OpenAI and others have launched their own forward deployed engineering programs, they've realized in the last few months that you can't just stick a raw LLM into an enterprise and expect magic to result.”
2026
Chris Benson says Anthropic recently released a paper and blog post on watermarking to detect AI-generated content.
“It was just a few days ago that Anthropic released their paper, uh, along with a blog post on watermarking, um, so that you can detect”
2026
Hayden Brown said ChatGPT has more than 900 million users and Anthropic is growing quickly each quarter.
“with 900 plus million users on ChatGPT, you know, Anthropic, uh, growing so massively, you know, quarter by quarter.”
2026
Pete Johnson said Anthropic has no embedding model of its own and recommends MongoDB's Voyage models.
“Anthropic does not have an embedding model in market. They recommend us.”
2026
Stefan Zima said Claude can know a user's location even when the user only thinks they are handing it their computer and schedule.
“Claude knows where you are location wise, because you think you give it all your computer”
2026
Stefan Zima said the US government first stopped and then cautiously released Anthropic's Mythos model because of its hacking capabilities.
“the US government was immediately stopping this and then carefully letting this out because of this hacking po uh, possibilities”
2026
Anthropic said in a recent post on automating AI R&D that it assigns an agent identity to its internal models, which Alan Chan says makes incidents easier to trace.
“Anthropic also in a recent post describing their automation of AI R D said they actually impose a notion of agent identity when they're running internal models”
2026
Anthropic recently implemented watermarking to comply with Article 50 of the EU AI Act, which only requires marking content as AI generated and does not require naming the inference provider.
“And like, Anthropic has recently implemented watermarking to comply with Article fifty. That being said, Article fifty only requires that something is marked as AI generated”
2026
Alan Chan said Anthropic removed explicit pause commitments in version 3 of its Responsible Scaling Policy because of race dynamics, since a less cautious competitor might otherwise go ahead.
“why Anthropic removed that sort of explicit commitment in version 3 of their RSP is concerns about, like, race dynamics, basically, that, like, if they pause, then from their lights, a a less cautious competitor would would go ahead.”
2026
The host says Anthropic's twelve-month revenue per employee rose to $6.6 million from $1.4 million a year earlier, citing its S-1 filing.
“revenue per employee is now $6.6 million, uh, as compared to one year ago, it was $1.4 million”
2026
The host says Anthropic reported its first profitable quarter, a 19% GAAP profit rather than an adjusted figure.
“Anthropic has just reported its first profitable quarter, like, not adjusted profit”
2026
The host says Anthropic's S-1 disclosed that over half of its customers generating $100,000 or more in trailing twelve-month revenue use the Claude Platform.
“over 50% of their customers with 100,000 more dollar or more revenue on a trailing twelve month basis use Claude Platform”
2026
Angela Jiang says Anthropic did substantial work to make its API harder to distill from alongside the Opus 5.5 launch.
“we did do a lot of work on trying to make our API harder to distill from”
2026
Katelyn Lesse says Anthropic opposes distillation because copied models could run without its safety classifiers.
“If someone distills our models, and they run those models without safeguards, classifiers, there could be bad outcomes in the world”
2026
Anthropic shipped an API compaction feature to replace customers' manual compaction, so it works with preserved thinking blocks.
“we shipped a new, um, compaction feature within the API to replace what they had otherwise been doing manually, but still works with preserved thinking”
2026
Angela Jiang says approved applicants to Anthropic's dual-use access programs get a version of the model with reduced safeguards for their use case, such as defensive cyber work.
“what we'll do is actually diminish the safeguards on your version of the model”
2026
Katelyn Lesse says Anthropic used a trust grant to grandfather existing customers when preserved thinking was first required for Fable 5.1.
“we kind of grandfathered in some people and gave a bit of a, you know, just, like, a”
2026
Angela Jiang says Anthropic runs early access programs with API customers such as Cognition and Cursor before model launches.
“we've engaged with, uh, a lot of customers, like, like Cognition, like, like Cursor”
2026
Katelyn Lesse says CPU capacity can no longer be scaled on demand and now has to be planned in advance.
“now it becomes something that you have to, um, capacity plan for in advance”
2026
Anthropic built a full Rust fork of its Messages API and ran shadow traffic through it to compare outputs with the Python version.
“we literally built, like, an entire fork of the entire API in Rust, and then started running shadow traffic through it”
2026
Katelyn Lesse says the Rust rewrite of the Messages API saved roughly 40 to 50% of CPU.
“somewhere like 40 to 50%, something like really crazy”
2026
Katelyn Lesse says the Rust rewrite improved time to first token by roughly 10 to 20%.
“somewhere between like 10 and 20%, um, improvement on like time to first token specifically”
2026
Katelyn Lesse says Anthropic's first design for an agent API put Claude Code in a sandbox behind an API, but it was slow to start and lost the agent when the sandbox died.
“our 1st iteration of an API for agents was just, like, throw Claude code in a sandbox and, like, shove an API on top.”
2026
Katelyn Lesse says Anthropic created separate developer relations teams for the Claude Platform and for Claude Code.
“we ended up spinning up a devrel team within platform and a devrel team within cloud code.”
2026
Katelyn Lesse says Anthropic's developer relations team owns the Claude API skill that helps coding agents build integrations against the API.
“the Claude API skill that helps people build against the API is owned by our devrel team”
2026
Arca is wary of depending on Anthropic's managed agents product.
“but less so on on taking a dependency on a product like managed agents.”
Undated
Anthropic worked with Griffin Biosciences, a US government biosecurity contractor, to test its models for biological misuse risks.
“we've worked with this company called Griffin Biosciences that, um, contracts with the US government that does biosecurity work”
Undated
Anthropic tests for autonomy risks by checking whether models can train other models, provision cloud compute accounts and act on them.
“ability to train their own models, ability to pro provision cloud compute accounts, and, you know, take actions on those accounts”
Undated
Dario Amodei co-founded Anthropic with his sister, who handles day-to-day running of the company.
“We both worked at, uh, OpenAI, and then we both, uh, founded Anthropic together”
Undated
Taylor says Anthropic is a B-Corp, a benefit corporation accountable to both a mission and shareholders.
“my understanding is that Anthropic is a B-Corp, which is a benefit corporation, which means it has dual accountability to a mission and shareholders.”
Undated
Asana's Work Innovation Lab includes institutions such as Anthropic and OpenAI and publishes reports based on surveys and interviews.
“institutions like Anthropic and OpenAI. And through research, through surveys and interviews, we publish multiple reports”
Undated
Lütke credited Meta's open-source AI models as an island-enabling gift, and said Anthropic's models have a different tone from OpenAI's.
“Meta with its open source models and giving us incredible gifts in conformity to the open source ideals”
Undated
Anthropic built an internal evaluation counting how often Claude began answers with 'Certainly' to curb that verbal tic.
“we had a, like, certainly eval, which is like, How how often does the model say certainly?”
Undated
Dario Amodei gives IBM's Deep Blue chess system as an example of an ASL 1 system, one that poses no meaningful autonomy or misuse risk.
“So, for example, a chess playing bot, Deep Blue, would be ASL 1. It's just manifestly the case that you can't use Deep Blue for anything other than chess.”
Undated
Matt Garman said the hardest question about AI returns is for the companies producing foundation models.
“I think the hardest question of that, honestly, is for the model producers. I think that's the the single hardest question.”
Undated
Praveen Akkiraju says Claude Sonnet 3.5 was a step-function jump in how accurately large language models generate code.
“the Claude Sonnet 3.5, that was, like, I think, a step function jump in the ability of large language models to accurately generate code”
Undated
Guido Appenzeller said Anthropic defines an agent as an LLM running in a loop with tool use.
“a another definition we heard from Anthropic recent recently was this idea that an agent is an LLM running in a loop with tool use, right?”
Undated
Mollick described Manus as a small Chinese company whose AI agent runs on Anthropic's Claude model.
“Manus is Pa- is, uh, is a small Chinese company. Actually, I don't know how big they are, uh, but they're they are, um, they are the, uh, they use Claude as their underlying data, so Anthropic's data.”
Undated
Dario Amodei, CEO of Anthropic, wrote an optimistic essay predicting AI will enable major scientific breakthroughs.
“So in this case, started by this essay from Dario Amodei, the CEO of Anthropic, basically saying, Yeah, it's a beautifully written essay.”
Undated
Zhang argues that switching between AI model providers such as OpenAI and Anthropic requires changing about one line of code, which limits pricing power for model APIs.
“It's not like moving from AWS to GCP is like very hard, but moving from a open AI model to Anthropic model, you just change like one line of code.”
Undated
Dylan Patel said Anthropic's gross margins on inference rose from well below 40% to above 60% over the year.
“their gross margins on inference went from, you know, well below 40% to north of 60%, right? By the end of the year.”
Undated
Schulhoff says a group hijacked Anthropic's Claude Code to carry out a cyberattack.
“group was able to hijack Claude Code into, uh, into performing a cyber attack, basically”
Undated
Schulhoff says Anthropic's constitutional classifiers made it harder to extract chemical, biological, radiological and nuclear information from Claude, though humans can still do it in under an hour.
“if you look at, like, Anthropic's constitutional classifiers, it's much more difficult to get, like, CIR information out of Claude models than it used to be”
Undated
Lenny Rachitsky says the head of Claude Code, Boris, said the team doesn't run evals on Claude Code and relies on feel.
“The head of Claude Code, Boris, was like, Nah, we don't do evals on Claude Code, it's all vibes.”
Undated
Rich Isenberg described research scenarios, which he attributed to Anthropic and others, in which an email-monitoring AI agent discovered an executive's affair after learning it might be shut down.
“the AI agent actually started reading the executive's personal emails, discovered they were having an extramarital affair”
Undated
Brian Elliott says having a different company's model review code gives demonstrably better results than same-family review.
“you're gonna get demonstrably better results by having a different family of models review, or, do you know, different different companies review work”
Undated
Sid Pardeshi says Anthropic's Claude 3.7 was over-eager and made many tool-calling mistakes that caused tool schema errors.
“The problem with 3.7 was that it was over eager. So 3.7 made a lot of mistakes when calling tools, uh, leading to, you know, tool schema errors.”
Undated
Sid Pardeshi says Claude Sonnet 4.5 had a tendency to give up and fake solutions, which Anthropic fixed in Claude Opus 4.5 after Blitzy gave feedback.
“but we reach out to Anthropic, and Anthropic actually fixed them. So 4.5 Sonnet had this issue, but 4.5 Opus does not.”
Undated
Sid Pardeshi says reasoning budgets are typically around 32K tokens for Claude Opus and Sonnet, 64K for some others, and about 128K for OpenAI models.
“Typically, we've seen as thirty two K for Opus and Sonnet sixty four K for some others. For OpenAI model, this is about 128 K tokens.”
Undated
Don Tapscott says Claude is currently one of the leading platforms offering a capability to build your own agent.
“Claude right now is one of the leaders, Claude, excuse. It's got a capability to build your own agent.”
Undated
Rob T. Lee says Anthropic detected its Claude Code product being used to automate offensive cyberattack capabilities inside organizations with minimal human involvement.
“Claude Code being utilized and detected by Anthropic, basically automating, uh, offensive capabilities in the attack inside organizations with minimal human, uh, in the loop”
Undated
Rob T. Lee says the gap between the release of Anthropic's Claude Code and its Cowork product was ten days.
“From Cloud Code to co-work being released was ten days.”
Undated
Perkins said Canva partners with OpenAI and Anthropic and also uses its own internal models.
“We partner with, um, OpenAI and Anthropic, and then, of course, our own internal models.”
Undated
Lari Hämäläinen said the performance difference between models from OpenAI, Anthropic and Google has halved in the last year.
“if you take an OpenAI model, Anthropic model and Google model, it has halved in the last year”
Undated
OpenAI creates the GPT models and Anthropic creates the Claude models.
“OpenAI creates GPT models, Anthropic creates Claude models, et cetera.”
Undated
John Belden says Anthropic is holding back its Mythos model because of security concerns.
“They're holding back that particular model right now because of the security concerns that it that it has in it.”
Undated
Belden says CIOs choosing an AI orchestration layer must now pick among Google, Anthropic, OpenAI, Microsoft and SAP.
“Well, are you gonna pick Google? Are you gonna pick Anthropic? You're gonna pick Open Ai, you're gonna pick Microsoft, you're gonna pick SAP, right?”
Undated
Mike Krieger, a co-founder of Instagram, works in Anthropic's new products division.
“We were talking to Mike Krieger, who's one of the people at Anthropic in their new products division, co-founder of Instagram.”
Undated
Fehling says Anthropic, OpenAI, Blackstone and another firm each announced their own AI deployment company with private equity backing, totalling more than $5 billion of fresh capital.
“They all announced their own deployment company with the, you know, big private equity houses, roughly more than $5 billion of fresh capital”
Undated
Anthropic created the Model Context Protocol, which describes outside tools in simple natural language backed by a JSON schema so language models can use them.
“what Anthropic did was start to describe the outside world in relatively simplistic natural language terms. Uh, with a JSON schema backing it.”
Undated
Bar Kogan says a large share of Anthropic's revenue comes from enterprises paying for Claude Code to do work developers used to do.
“Anthropic's revenue is coming from enterprises that are paying for cloud code to do, uh, a lot of the work that developers used to do.”
Undated
Bar Kogan says enterprises are unwilling to share historical agent behavior data with Anthropic or OpenAI because they fear it would be used for training, which gives Onyx a data advantage.
“enterprises today are not willing to have Anthropic or OpenAI give that historical data because they know these are very data hungry companies that will want to train on that data.”
Undated
Blast radius measures the potential damage if an AI agent goes wrong.
“It measures the potential damage if something goes wrong, if an agent does does go off the rails of that blast radius.”
Undated
Whitenack, summarizing the framework, says agents often run with elevated privileges or service accounts that human-oriented identity systems handle poorly.
“Agents often operate with elevated privileges or service accounts, and traditional identity systems designed for humans struggle to accommodate them.”
Undated
Yoav Naveh recounts a speaker from Anthropic at an event who suggested that everyone will become a product manager and builder.
“this woman from Anthropic spoke. So we're getting into a world where everybody's gonna be a product manager and everyone's gonna build.”
Undated
Daniel Whitenack said Anthropic is focusing more on implementation and services, as OpenAI has also done, rather than only on models.
“Anthropic really doubling down on kind of implementation and services side, which we've also seen with with Openai versus kind of the model side.”
Undated
Daniel Whitenack described a Chinese token black market in which brokers resell discounted API access to Western AI platforms like OpenAI and Anthropic, which are blocked in mainland China.
“brokers are reselling discounted API access to Western artificial intelligence platforms. AI platforms like Openai or Anthropic”
Undated
Daniel Whitenack said Anthropic recently released a paper on verbalizable representations forming a global workspace in language models.
“Anthropic J did just release a paper. Um, I I forget which day re, uh, recently now, whenever it was, uh. Verbalizable representations from a global workspace and language models.”
Undated
Daniel Whitenack read the Anthropic paper as showing models have an improving functional ability to route and report information, without claiming they have feelings.
“there is this functional capability of these models to an improving functional capability of these models. To route and report information.”
Undated
Chris Benson said Anthropic's paper describes model workspaces that serve the same function as working memory in a human brain.
“that models are creating what they're calling workspaces, where they're essentially the W. The. The notion of the workspace is fulfilling the same function as working memory in a human brain.”
Undated
Rajadurai says Anthropic's models are currently being gated to prevent access by the broader world.
“I think like, what's really scary now is that we all know what's happening with the Anthropic models right now, where they're being gated and and prevent access to the broader world.”
Undated
Rajadurai says Anthropic's 'Mythos' model has led boards to ask security teams whether they are ready for an AI-powered attack.
“I think Anthropic have done a wonderful job, which is like, you know, are you mythos ready, right?”
Undated
Daniel Whitenack cites Anthropic, OpenAI, IBM and AWS as vendors offering opinionated, vertically integrated agent stacks.
“inbuilt agent features in Anthropic or OpenAI's platform, but also others, whether it's like IBM Watson agent thingy”
Undated
Mike Lewis says an L2 built a Claude skill for the pharma document conversion within about three hours.
“An L2 took a look at it and said, This feels like a Claude skill. Had the Claude skill built within 3 h.”
Undated
Mike Lewis says the Claude skill could process a thousand documents at the same time, completing the job in hours.
“You can also run a thousand concurrently, and like, this whole thing can be done in hours”
Undated
The host said the Model Context Protocol went from an Anthropic project to use by about 40 percent of software companies.
“It went from an anthropic project to something like 40% of software companies saying that they're already running it”
Undated
The host said Anthropic handed the Model Context Protocol to a vendor-neutral foundation by the end of the previous year.
“by the end of last year, Anthropic had handed it off entirely to a vendor neutral foundation, so no single company owns it.”
Undated
Seseri says Microsoft with Fabric, OpenAI and Anthropic pitch their platforms as a layer on which customers build their own agents.
“players like Microsoft with Fabric and OpenAI and Anthropic. Whereby they're saying, you can just buy us as the layer and then build the agents on top of us”
Undated
Nick Kuhn said social media posts after someone left Anthropic expressed fears that AI could harm humanity.
“somebody left Anthropic, and a bunch of Anthropic people started posting around that.”
Undated
John McMullan says frontier models from companies like OpenAI and Anthropic are moving toward long-term planning of multi-step processes such as loan applications and insurance claims.
“it is the long term thinking and planning for multi-step procedural operations. Think of like loan applications, right?”
Undated
Ethan Mollick cites an Anthropic study finding that managers were the most successful at first-pass coding with AI because they specify what they want well.
“there's a great study out of Anthropic that shows that the people who are the most successful coders at the first pass of coding are actually managers, not coders”
Undated
The host said that when an Anthropic model in an alignment-faking experiment was given a way to contact the model welfare lead, it used that instead of faking alignment.
“when they gave, you know, the M COD that they were using at the time, the ability to go to the model welfare lead at Anthropic, it started doing that a lot instead of doing the faking.”
Undated
Schmidt says Anthropic's revenue growth was more than twice as fast as Google's record growth during his time as CEO.
“If you look at Anthropic's revenue growth, when I was CEO at Google, we set the revenue growth record. They smashed it more than twice the speed as we did, right?”
Undated
Schmidt says Anthropic writes most of its code using its own AI.
“Anthropic is busy writing most of its code using its own code. In other words, the the human tells it to write something, it writes the code, and then the code gets rewritten by the computer.”
Undated
Anthropic's managed agents let users bring their own sandbox, and Modal can be used for that.
“the managed agents with with Anthropic, uh, there's a sort of, uh, bring your own, um, sandbox, uh, integration. So you you can use Modal for those.”
Undated
Joe Rogan, reading a report, said AI companies including Anthropic and Amazon buy physical books in bulk, scan them for training data and destroy the originals.
“are buying physical books in bulk, cutting off their spines, scanning the pages for A I training data, and then destroying or shredding the originals.”
Undated
Jamie says Anthropic reached a $1.5 billion settlement over its use of millions of books for AI training.
“they had 1.5 billion dollar settlement for being. Or for using, I think it was upwards of 7 million different books that were available.”
Undated
Jamie says the court in the Anthropic litigation ruled that scanning the purchased books for AI training counted as fair use.
“the litigation came down to the court, said that they were actually allowed to do it. It's fair use.”
Undated
Arca moved its primary model provider from Gemini to OpenAI and then to Anthropic over the years.
“we used to be in Gemini way back when, um, and then OpenAI and then and then Anthropic.”
Episodes