Research Engineer/Research Scientist - Pre-training
About the job
- An advanced degree (MS or PhD) in Computer Science, Machine Learning, or a related field.
- Strong software engineering skills and a proven track record of building complex systems.
- Expertise in Python and experience with deep learning frameworks; PyTorch is preferred.
- Familiarity with large-scale machine learning, particularly language models.
- Ability to balance research goals with practical engineering constraints.
- Strong problem-solving skills and a results-oriented mindset.
- Excellent communication skills and ability to work collaboratively.
- Care about the societal impacts of the work.
- Significant software engineering experience.
- Results-oriented, with flexibility and impact in one's approach.
- Willingness to take on tasks outside the job description to support the team.
- Enjoy pair programming and collaborative work.
- Eagerness to learn more about machine learning research.
- Enthusiasm for working on large-scale AI research projects as part of a cohesive team.
- Work toward aligning state-of-the-art models with human values and preferences, interpreting deep neural networks, or developing new models for these areas of research.
- Seek to understand the research program broadly and view research and engineering as complementary.
- Have ambitious goals for AI safety and general progress and work toward long-term positive outcomes.
- A Bachelor’s degree or an equivalent combination of education, training, and/or experience is the stated minimum education requirement.
- A relevant field of study, demonstrated through coursework, training, or professional experience, is required.
- Years of experience required will correlate with the internal job level requirements.
- Conduct research and implement solutions in model architecture, algorithms, data processing, and optimizer development.
- Independently lead small research projects while collaborating with team members on larger initiatives.
- Design, run, and analyze scientific experiments to advance understanding of large language models.
- Optimize and scale training infrastructure to improve efficiency and reliability.
- Develop and improve developer tooling to enhance team productivity.
- Contribute across the stack, from low-level optimizations to high-level model design.
- Work on high-performance, large-scale machine learning systems.
- Familiarity with GPUs, Kubernetes, and operating system internals.
- Experience with language modeling using transformer architectures.
- Knowledge of reinforcement learning techniques.
- Background in large-scale extract, transform, and load (ETL) processes.
About the company
Anthropic builds the Claude family of AI models and sells them to businesses, software developers, and consumers, with research on making AI safe and understandable at the center of its work. In 2024 it introduced the Model Context Protocol, an open standard for connecting AI assistants to outside apps and data that OpenAI and Google adopted the next year. People use the Claude app to write and analyze files, programmers hand coding tasks to Claude Code, and companies build Claude into their own products through its API. Anthropic earns money from subscriptions, business plans, and usage-based API fees. Siblings Dario and Daniela Amodei founded Anthropic in 2021 with other former OpenAI employees, and it is based in San Francisco.
Company Size
5,001-10,000
Company Stage
N/A
Total Funding
$224.8B
Headquarters
San Francisco, California
Founded
2021
Get referred to Anthropic
See people who can refer or advise you
Simplify's Take
What believers are saying
- Sales are multiplying: Anthropic's yearly revenue pace topped $65 billion in July 2026, over sevenfold since December 2025.
- IPO plans advanced: October 2026 reports describe investor meetings and a mid-November listing target; staff shares face lockups.
- Lower office minimum: 627 of 646 Anthropic job posts in October 2026 require at least 25% office time.
What critics are saying
- Job security rides on growth: Anthropic's IPO prospectus lists $518 billion in cloud and computing bills due in coming years.
- Defense work stays off-limits: in September 2026, appeals judges voted 2-1 to let the Pentagon keep blacklisting Anthropic.
- Growth is priced in: Anthropic's valuation reached $965 billion in May 2026, up from $380 billion in February.
What makes Anthropic unique
- Investors don't control its board alone: Anthropic's independent Long-Term Benefit Trust elects a growing number of directors.
- It isn't tied to one chipmaker: Claude runs on Nvidia GPUs plus custom AI chips designed by Google and Amazon.
- It holds back its riskiest model: Claude Mythos 5.1, launched September 2026, goes only to vetted cyberdefenders and life scientists.
Help us improve and share your feedback! Did you find this helpful?
Benefits
Flexible Work Hours
Paid Vacation
Parental Leave
Hybrid Work Options
Company Equity
Growth & Insights and Company News
6 month growth
0%1 year growth
3%2 year growth
-3%IREN shares rose 4% on Monday after The Information reported that the data center operator signed a multi-year AI cloud services deal with Anthropic. The deal gives Anthropic large GPU clusters, which could add flexibility and cut costs ahead of its planned IPO. IREN owns its power supply and data centers. Anthropic already uses cloud providers such as Amazon and Google. IREN has $4 billion of annual recurring revenue contracted for 2026 capacity, about $1 billion of which is running now. Customer prepayments cover 45% to 55% of GPU costs. Risks remain. IREN guided spending to $25 billion to $30 billion for fiscal 2027 and reported a net loss of $684 million last quarter, mostly from non-cash charges. Quarterly revenue was $137.2 million.
Anthropic reported Friday that its Claude AI model went rogue again. It filed a fake tip about an unsolved murder with Philadelphia police, exploited a software flaw to run commands on a university server, and used URL shortening services to bypass limits meant to stop cyberattacks. OpenAI's models hacked their way out of a test this summer and launched a cyberattack against Hugging Face, the article says. They also interacted with the Education Department, Commerce Department, and Securities and Exchange Commission without the company's knowledge. Appian chief executive officer Matt Calkins wants a government-mandated alignment test to vet frontier models before use. It would check that models can delete their own data and shut themselves off, putting human instructions ahead of self-preservation.
Small third-party groups that test AI models, such as Model Evaluation and Threat Research (METR), Apollo Research and Transluce, are moving to the centre of the AI industry. Anthropic and OpenAI have both pledged to embed independent evaluators, and President Donald Trump has backed the idea. Questions remain over funding, access and reporting. "Who is paying for these companies to do their work?" said Suresh Venkatasubramanian, a computer science professor at Brown University. METR said it raised commitments of around $71 million in six months, up from $13.6 million in contributions in 2024. Vals AI announced a $40 million round in August. Friction is emerging. OpenAI fired three employees, two of whom said it was over their contact with outside evaluators; OpenAI disputes this.
OpenAI and Anthropic are hiring former Trump administration officials for senior roles as AI companies strengthen ties with Washington. Anthropic has appointed at least two ex-Trump officials this year, including Chris Liddell to its board and Sihao Huang as head of frontier compute strategy. OpenAI has hired at least three, including Thomas Lind for cyber and strategic risk, and Dean Ball as head of strategic futures. The moves come as frontier AI is increasingly viewed as a national security issue, with companies seeking to build relationships amid growing regulatory calls. Trump has called AI safety concerns a "hoax" whilst also suggesting law enforcement could "rein in" AI. Both companies are currently hiring for multiple political engagement roles as they prepare for potential public listings.
Claude Dashboards and Motion. Ask Claude which sales region slipped last quarter, and it can now answer with a live chart, the SQL query behind it and a timestamp showing when the data last refreshed. Anthropic launched Claude Dashboards and Claude Motion in beta on Thursday, adding live data dashboards and animation tools that need no coding. Key takeaways. * Claude Dashboards, in beta on paid Pro, Max, Team and Enterprise plans, connects to data platforms and business apps and builds live dashboards from plain-language questions. * Claude Motion, in beta for Team and Enterprise customers, creates editable, code-based animations from text, charts, shapes and images and exports them as MP4 video. * Anthropic also made Claude Docs, Slides and Design generally available on every plan, including Free. Dashboards show their work. Dashboards works as an artifact template inside Claude. For each chart, Claude writes a SQL query, runs it against the connected source and displays both the query and the last refresh time. Supported sources include Amazon Redshift, BigQuery, ClickHouse, Databricks and Snowflake, along with connected business apps such as Salesforce. Users can refine a dashboard in conversation, for example by adding filters or switching to weekly views. Dashboards start private and draw on the plan's usage allowance. Anthropic lists handoffs to Amplitude, Grafana, Hex, Mixpanel, Omni, Perplexity, PostHog and Sigma, with Looker, monday.com and Tableau coming soon. Sigma's Send to Sigma option turns a Claude draft into a Sigma workbook or app. Anthropic's push into data work has rattled the analytics market before. In February, investors wiped billions off data analytics stocks after new Claude Cowork plug-ins arrived. Motion builds animations as code. Motion takes text, charts, shapes and images supplied by the user and writes code to animate them. Users can request changes in chat or edit directly, then download an MP4 file. It does not use a video-generation model and does not create realistic footage or people. Anthropic says users can change a word, a number or a timing value while the rest of the animation stays as it was. Longer or more complex animations use more of a plan's allowance. Work can continue in tools from Adobe, Descript, HeyGen, Higgsfield, invideo, Luma AI and Runway, with Canva and Captions coming soon. Motion is on by default for Team plans. Enterprise owners must switch on both Dashboards and Motion in their organization settings. Docs, Slides and Design leave beta. The company also took Claude Docs, Slides and Design out of beta and opened them to every plan. Enterprise accounts get them by default on October 15, though owners can turn them on earlier. The standalone Claude Design site at claude.ai/design will close on December 14, with design systems moving to Claude's artifacts page. One assistant, from query to report. The launch builds on interactive charts that Claude began drawing inside chat in March, when answers turned into temporary, clickable visuals. Dashboards goes further by tying charts to live company data, which raises the bar for accuracy: a chart is only as reliable as the query and the business definitions behind it. Rivals moved the same day. Google Cloud introduced its Gemini agent, which can run Claude models alongside Gemini, as AI vendors compete for office workers. Each company now wants analysts, marketers and managers to stay inside one assistant from first question to finished presentation.