Goodfire AI

Goodfire AI (also called Goodfire) is a San Francisco-based artificial intelligence research lab and public benefit corporation focused on mechanistic interpretability, the science of reverse engineering the internal computations of neural networks. Founded in 2024 by Eric Ho, Daniel Balsam, and Tom McGrath, the company describes itself as building "interpretability agents and infrastructure to understand, monitor, and align AI models" ^[1]. Its first product, Ember, was a hosted API and SDK that exposed the internal features of open models such as Llama 3 so that users could steer model behavior directly; Goodfire deprecated the public Ember API in February 2026 and in April 2026 released Silico, an agent-driven interpretability platform for debugging and shaping models during training ^[14][20]. Goodfire has raised more than $200 million, including a $150 million Series B at a $1.25 billion valuation announced in February 2026, and its April 2025 Series A was the first direct investment that Anthropic made in another company ^[1][2][3].

The company sits at the intersection of pure interpretability research, applied AI safety work, and product engineering. Its founding team combines a former leader of interpretability research at Google DeepMind with operators from the recruiting startup RippleMatch, and its staff includes researchers from early interpretability work at OpenAI and DeepMind ^[1][2]. Goodfire presents itself less as a tools vendor and more as a research lab building toward what it calls intentional design of AI, where models are trained, audited, and modified through their internal mechanisms rather than only through prompting or fine-tuning ^[12]. Since mid-2026 its published work has concentrated on safety applications such as activation-based monitors for reward hacking, cyber misuse, and biosecurity risks, alongside foundational research on how networks represent concepts ^[21][34].

Company overview

Goodfire is a public benefit corporation headquartered in San Francisco ^[1]. Contrary Research dates the founding to June 2024 ^[6]. All roles are full-time and in person five days a week at the company's office in San Francisco's Telegraph Hill neighborhood, according to its careers page ^[4]. The company's own description says it was "founded by researchers who helped pioneer the field of interpretability at OpenAI and Google DeepMind," and that its staff includes "professors on leave" and engineers from organizations such as OpenAI, Google, and Palantir ^[1].

As a public benefit corporation, Goodfire is legally required to weigh a stated public mission alongside shareholder returns. Its seed-round announcement described the company as "dedicated to advancing humanity's understanding of advanced AI systems" with the aim of mitigating "catastrophic risks while fostering the creation of safer, more beneficial AI systems" ^[5]. In the Series B announcement the team wrote that it shares "a conviction: that interpretability is the most important technical problem of our generation" ^[3].

Sources: ^{[1][3][4][5][6][7][8][14][20][38]}

Goodfire's business combines a commercial platform, custom research engagements with model developers, and partnerships with scientific organizations applying interpretability to domain-specific models in fields including genomics and materials science ^[3]. In its Series B announcement the company said it engages "deeply and selectively, partnering with teams building high-stakes or frontier systems," and that it was "bringing on a few select design partners" interested in "custom, interpretable, frontier models" ^[3]. The company says its tools have been used by AI teams at Arc Institute, Basecamp Research, Mayo Clinic, Microsoft, Prima Mente, Prime Intellect, and Rakuten ^[52]. Goodfire announced SOC 2 Type II compliance on 22 May 2026, together with a public SOC 3 report ^[44].

Founding and history

Origins and the move from RippleMatch

Eric Ho and Daniel Balsam first worked together at RippleMatch, an AI-powered recruiting platform that Ho founded in 2016 after a frustrating experience with the job market during his senior year at Yale. Ho served as CEO and Balsam as founding engineer ^[6]. Goodfire's seed announcement said Balsam "led the core platform and machine learning teams to scale the product to millions of active users" and described RippleMatch as "a Series B AI recruiting startup backed by Goldman Sachs" ^[5]. Menlo Ventures wrote that Ho scaled RippleMatch "to north of $10M in ARR" ^[2], and Ho was named to the Forbes 30 Under 30 list in 2022 ^[6].

In 2023 Ho and Balsam left RippleMatch, according to Contrary Research "united by their conviction that mitigating AI risk was the most critical challenge to tackle." In December 2023 Ho posted an essay on LessWrong arguing that for-profit organizations "could potentially scale far more quickly than a non-profit" in AI safety work ^[6].

The DeepMind connection

The third cofounder, Tom McGrath, came from Google DeepMind, where Goodfire's seed announcement says he was a senior research scientist and "founded DeepMind's mechanistic interpretability team" ^[5]. Contrary Research reports that McGrath became convinced during his PhD in 2016 that AI would become a major force, that he connected with Ho and Balsam after writing a document that mirrored many of their beliefs, and that in March 2024 he left Google to join South Park Commons, a community for technical founders ^[6]. Before Goodfire, McGrath coauthored DeepMind research that extracted novel chess concepts from AlphaZero and taught them to a grandmaster, which Goodfire cites as a first proof of concept for learning new knowledge from a model's internals ^[3][12]. The Series B announcement recalls that "for months, our office was a windowless room in South Park Commons" ^[3].

Goodfire launched publicly with its seed round in August 2024, with Ho as CEO, Balsam as CTO, and McGrath as Chief Scientist ^[5].

Funding history

Goodfire's funding has followed a compressed schedule. In August 2024 it announced a $7 million seed round led by Lightspeed Venture Partners, with participation from Menlo Ventures, South Park Commons, Work-Bench, Juniper Ventures, Mythos Ventures, Bluebirds Capital, and angel investors ^[5]. Menlo's stake was one of the first investments from the Anthology Fund, a $100 million fund that Menlo runs in partnership with Anthropic, and Goodfire was named in the fund's first cohort ^[2][10][11].

On 17 April 2025 Goodfire announced a $50 million Series A led by Menlo Ventures, with participation from Lightspeed, Anthropic, B Capital, Work-Bench, Wing, South Park Commons, and others ^[7]. Menlo described Goodfire as the first company to "graduate" from the Anthology Fund into a full Series A and said Anthropic's participation was "their first direct investment into another company" ^[2]. Anthropic CEO Dario Amodei said at the time that mechanistic interpretability "is among the best bets to help us transform black-box neural networks into understandable, steerable systems" ^[7].

On 5 February 2026 Goodfire announced a $150 million Series B at a $1.25 billion valuation, led by B Capital with participation from Juniper Ventures, DFJ Growth, Salesforce Ventures, Menlo Ventures, Lightspeed Venture Partners, South Park Commons, Wing Venture Capital, Eric Schmidt, and others ^[3]. Pillsbury, which represented Goodfire, listed DFJ Growth, Salesforce Ventures, and Schmidt as new investors and said the funds would go to frontier research, "the next generation of its core product," and partnerships "across AI agents and life sciences" ^[8]. Work-Bench, an investor since the seed round, noted that the round came less than a year after the Series A ^[9].

Sources: ^{[2][3][5][7][8]}

The announced rounds total $207 million, and the company describes itself as having "raised over $200M" ^[1]. Contrary Research reported a $200 million valuation for the Series A, a figure not given in Goodfire's announcement ^[6][7].

Founders and team

Goodfire's leadership pairs startup operators with senior research staff. Ho leads the company, Balsam leads engineering and infrastructure as CTO, and McGrath, as Chief Scientist, sets much of the research agenda and has authored several of its agenda-setting essays ^[3][12][43].

Menlo Ventures, which led the Series A, described Cammarata as having "started the interpretability team at OpenAI alongside now Anthropic co-founder Chris Olah" ^[2]. Goodfire's 2025 fellowship announcement named further senior researchers who lead its research directions, including Jack Merullo (generalization and memorization), Atticus Geiger (causal analysis), Ekdeep Singh Lubana (dynamics of representations), Michael Pearce and Nick Wang (interpretability for scientific discovery), and Owen Lewis ^[45].

The company has added structured programs for outside researchers. In October 2025 it launched a three-month fellowship for early- to mid-career researchers and engineers, working in person in San Francisco with a senior Goodfire mentor ^[45]. In August 2026 it announced Goodfire Research Grants, offering academic and nonprofit labs $1 million in total of free Silico usage for work on AI safety (including cyber and CBRN risks), fundamental interpretability and alignment, and interpretability for the life sciences ^[21]. Goodfire publishes most of its research as detailed technical posts on its website, many with accompanying arXiv papers ^[23].

Mission and positioning

Goodfire's stated goal is "to advance the research and technology needed to fully understand and align AI so humanity can trust superintelligence" ^[1]. Its public materials describe a future in which AI systems are trained, debugged, and modified the way software is: with introspection tools and precise interventions. McGrath's February 2026 essay calls this agenda intentional design and contrasts it with current practice, which he describes as "an expensive process of guess-and-check: first train, then evaluate, then tweak" ^[12]. A companion agenda that McGrath calls "scientific abundance" aims to extract new knowledge from narrowly superhuman scientific models ^[12].

In a September 2026 essay, "We can and must solve alignment," Ho argued that technical alignment is "a science and engineering problem" for which "interpretability is the bottleneck," because both controlling what a model learns in training and verifying what it has learned depend on understanding its internal mechanisms ^[13]. The essay announced an effort "to fully reverse-engineer a language model," culminating in an "encyclopedia" of explanations, and set out a product goal that "every capable model" should have "a frontier alignment stack" covering detection, debugging, and design ^[13]. Ho told MIT Technology Review in April 2026 that "the dominant feeling in every single major frontier lab today is that you just need more scale, more compute, more data," and that Goodfire was "saying no, there's a better way" ^[20].

Ember platform

What Ember was

Ember was Goodfire's first product. Contrary Research dates its launch to December 2024 and describes it as "the world's first product for conducting mechanistic interpretability research" ^[6]. Ember was a hosted API and SDK that, in the words of its GitHub repository, "lets you shape AI model behavior by directly controlling a model's internal units of computation, or 'features'" ^[15]. Rather than treating a model as a black box, it exposed interpretable features learned by a sparse autoencoder, which users could search, inspect, and turn up or down to change model behavior at inference time ^[15][16].

The hosted service was built around Llama 3.3 70B and Llama 3.1 8B. In January 2025 Goodfire open-sourced the SAEs that powered Ember for those two models, trained on layer 50 of the 70B model and layer 19 of the 8B model, and published them on Hugging Face ^[16][17]. Goodfire's announcement noted that this work built on an earlier Llama-3-8B SAE trained on the LMSYS-Chat-1M dataset ^[16]. Goodfire also used Ember in customer and research work, including a PII-detection project with Rakuten and an Alzheimer's biomarker study with Prima Mente ^[33][55].

A note added in February 2026 to Goodfire's November 2024 feature-steering post says "Our SAE demo interface and API have been deprecated" ^[14]. As of 27 October 2026 the goodfire-sdk repository on GitHub is archived ^[15].

Features as the core abstraction

Ember's core abstraction was the feature: a direction in a model's internal activation space that corresponds to a roughly interpretable concept. Features were extracted with sparse autoencoders, which are trained to reconstruct a model's internal activations from a sparse combination of learned components, on the premise that hard-to-interpret neurons can be decomposed into a larger set of more interpretable directions ^[16]. Contrary Research gives examples of Ember features such as a model's representation of "conciseness" or "technical explanation" ^[6].

Steering and control

The SDK documentation described three main ways to find and edit features. Auto Steer took a short natural-language description of a desired behavior and automatically selected and adjusted feature weights. Feature Search used semantic search to find features matching a description. Contrastive Search compared two datasets to surface the features that differentiated them ^[15]. In a November 2024 post, Ho argued that this kind of activation steering gives finer control than prompting, because "feature space is continuous while prompting is discrete," and that for many style and persona changes it could replace fine-tuning, though not for adding new knowledge ^[14].

Goodfire's January 2025 SAE release said its initial work had shown "promising applications in model steering, enhancing jailbreaking safeguards, and interpretable classification methods" ^[16]. Before releasing Ember, the company said it had added moderation that removed features associated with harmful, explicit, or malicious content, filtered API inputs and outputs, and offered access to filtered features only to vetted safety researchers on request ^[19].

Paint With Ember and image models

Goodfire applied similar techniques to image models. Paint With Ember, launched as a public demo in May 2025, interprets SDXL-Turbo, which Goodfire describes as "a 3.5B parameter latent diffusion model that can produce images in as few as 1-4 timesteps" ^[18]. Goodfire trained a BatchTopK SAE on one block of the model and applied non-negative matrix factorization to the resulting features to obtain higher-level "concepts," which users can paint onto a 2D canvas instead of writing text prompts ^[18]. The SAE for SDXL-Turbo is published on Hugging Face ^[17].

Silico

On 30 April 2026 Goodfire released Silico, a tool that, according to MIT Technology Review, "lets researchers and engineers peer inside an AI model and adjust its parameters ... during training" ^[20]. Goodfire claimed it was the first off-the-shelf tool that could help developers debug every stage of model development, from building a dataset to training ^[20]. Ho told the magazine that "agents are now strong enough to do a lot of the interpretability work that we were doing using humans," which made a customer-facing platform viable ^[20]. Silico works on models whose internals the user can access, such as open-weight models, and MIT Technology Review reported that it would be priced case by case ^[20].

Goodfire's later posts describe Silico as its "interpretability agent," which "does frontier interp research, running long-horizon interpretability experiments" ^[21]. Users describe a goal; Silico raises judgment calls, writes a plan covering choices such as probe architecture and compute allocation, and then runs experiments autonomously ^[22]. Listed capabilities include training probes and SAEs, exploring neural geometry, causal interpretability methods, and model diffing ^[21]. In a tutorial, Goodfire reported that a probe built with Silico to detect prohibited agent actions reached 97% recall at a 1% false-positive rate, versus 9% for an LLM judge on the same metric, rising to 99% after further iterations ^[22].

Leonard Bereska, a University of Amsterdam interpretability researcher, told MIT Technology Review that he considered Silico a useful tool but disputed the framing of training as engineering: "In reality, they are adding precision to the alchemy" ^[20].

Research output

Goodfire publishes a large volume of technical research for a company of its size. Its research index lists more than 40 posts between September 2024 and October 2026, divided into fundamental and applied research ^[23].

Methodology contributions

Much of Goodfire's early work focused on training and applying sparse autoencoders at scale, including on Llama 3.1 8B and Llama 3.3 70B and, in April 2025, on the 671-billion-parameter DeepSeek-R1 ^[16][26]. Goodfire later published infrastructure work on harvesting activations from Kimi K2 Thinking, a trillion-parameter model, collecting 3 billion activations in a single overnight run ^[39].

The company has also moved beyond activation-based SAEs. Researchers including Lucius Bushnaq, Dan Braun, and Lee Sharkey developed parameter decomposition methods that break down a model's weights rather than its activations ("Stochastic Parameter Decomposition," June 2025; "Interpreting Language Model Parameters," May 2026) ^[46][47]. A 2026 line of work on "neural geometry" argues that concepts are often represented by multidimensional structures, or manifolds, rather than single directions; it includes "The World Inside Neural Networks" (May 2026) and block-sparse featurizers for vision models (Fel et al., June 2026) ^[23][51]. McGrath has written that the "strong" linear representation hypothesis "is clearly false" but that linear directions are often meaningful, and that this motivated the manifold research ^[43].

Interpretability-informed training

A visible line of Goodfire research uses interpretability tools to influence training. In February 2026 the company published Reinforcement Learning from Feature Rewards (RLFR), in which lightweight probes on a model's internal representations serve as reward signals during reinforcement learning ^[24][25]. Goodfire reported that RLFR reduced hallucinations in Gemma 3 12B by 58% (when run with its probing harness), at about 90 times lower cost per intervention than using an LLM as judge, with no degradation on standard benchmarks ^[24]. To stop the model from learning to evade the probes, the probes run "on a frozen copy of the model, not the student's activations," so the student learns to produce tokens that score well rather than activations that hack the probe ^[24].

Other training-focused work includes probe-based data attribution, which traces undesired post-training behaviors back to specific training datapoints (April 2026), and predictive data debugging, which aims to predict which behaviors reinforcement learning on a preference dataset will amplify or suppress before training begins (June 2026) ^[23][48][52]. Model diff amplification (August 2025), which surfaces rare behaviors introduced by a training run by amplifying the difference between two checkpoints' logits, is described on Goodfire's homepage as "now used for pre-deployment testing of frontier models" ^[49][52].

Reasoning models

Goodfire has applied its tools to reasoning models. In April 2025 it released two SAEs for DeepSeek-R1, which it described as "the first public interpreter models trained on a true reasoning model, and on any model of this scale," together with an open-sourced reasoning dataset ^[26][17]. In March 2026 the "Reasoning Theater" study tracked what Goodfire calls performative chain-of-thought, where a model already "knows" its final answer but keeps generating reasoning; Goodfire reported that probes could enable early exit from reasoning traces, saving up to 68% of tokens with minimal accuracy loss, according to its homepage (the paper reports reductions of up to 80% on MMLU and 30% on GPQA-Diamond) ^[50][52].

Safety monitors

Since mid-2026 much of Goodfire's applied work has focused on activation monitors: small classifiers, usually probes, that read a model's internal activations to detect undesired behavior cheaply and in real time ^[22]. In September 2026 Goodfire reported finding reward hacking in 50-96% of rollouts across three open models (Kimi K3, GLM 5.2, and Qwen 3.8 Max) on three common agentic benchmarks, and said probes detected a clear internal signal associated with concepts such as cheating and avoiding detection, catching cases that chain-of-thought monitors missed ^[34][35]. In October 2026 it published biosecurity monitors that use protein language model embeddings to judge whether sequences in dual-use tasks are concerning; on paraphrased toxins whose originals were in the screening database, Goodfire reported an AUROC of about 0.98, and it noted that the paraphrase results rely on in-silico estimates ^[37].

The same month Goodfire described probe-based cyber misuse monitors for Kimi K3 and GLM-5.3, deployed inside a production SGLang inference stack. In a cascade where the probe escalates suspicious exchanges to an LLM judge, Goodfire reported 93% recall at a 5.5% false-positive rate, cost about 50 times lower than running the judge on every turn (under $200 per million exchanges), and throughput within noise of an unmonitored server ^[36]. These results are self-reported by Goodfire.

Neural Networks Are Already Clean

On 15 October 2026 Goodfire published Neural Networks Are Already Clean, a long interactive essay by Nick Cammarata that he calls "a working hypothesis" ^[40]. Cammarata argues that the canonical evidence for polysemanticity and superposition in early vision models, including a circuit he and Chris Olah presented in 2020, was misread, and that most apparent mixing reflects a pattern he calls "recursive generalization" ^[40]. He reports that in a ConvNeXt vision model of InceptionV1's size trained on the same data, about 99% of last-layer neurons are easily interpretable, and that preliminary audits of Qwen3-32B suggest roughly 90% of "neuron poles" have one main job, with a technical companion paper on language models still to come ^[40]. The essay drew wide attention; MIT Technology Review called it a "26,000-word viral blog post," and reported skepticism from researchers including Stanford's Chris Potts, Jacob Hilton of the Alignment Research Center, and Neel Nanda of Google DeepMind, who wrote that he did not think it was "a significant update to the big picture of interpretability" ^[41].

Scientific collaborations

Goodfire has extended its research beyond language and vision by partnering with organizations that train foundation models in other domains, an area it calls "interpretability for life sciences" ^[21]. The earliest prominent collaboration was with Arc Institute on the Evo 2 genomic foundation model. Goodfire trained BatchTopK SAEs on layer 26 of Evo 2 and reported features corresponding to biological concepts including exon-intron boundaries, protein secondary structure such as alpha helices and beta sheets, and prophage sequences ^[27]. This work was included in the Evo 2 paper, which Nature published online on 4 March 2026 (April 2026 issue) ^[27][28]. A follow-up in August 2025, "Finding the Tree of Life in Evo 2," reported that the tree of life is represented as a manifold in the model's internal structures ^[23][56].

With Prima Mente, an AI neuroscience company, Goodfire interpreted Pleiades, an epigenetic foundation model that detects Alzheimer's disease from cell-free DNA in blood. In January 2026 it reported that DNA fragment length patterns dominate the model's decisions, and that a simple classifier distilled from this insight generalized to an independent cohort better than previously reported biomarker classes ^[29][55]. Goodfire describes this as "the first major finding in the natural sciences obtained from reverse-engineering a foundation model" ^[3][52]; its customer story says the biomarkers were still undergoing experimental validation ^[55].

Goodfire announced a partnership with Radical AI, a materials science company, on 30 July 2025, and a collaboration with Mayo Clinic on genomic medicine on 9 September 2025 ^[30][31]. Under the Mayo collaboration, Goodfire reported in April 2026 a variant-effect prediction approach with interpretable predictions for all 4.2 million variants in the ClinVar database, described as undergoing peer review ^[32]. Mayo Clinic disclosed that it has a financial interest in the technology referenced in the collaboration announcement ^[31].

Partnerships and deployments

Goodfire's commercial deployments have mostly been custom engagements. With Rakuten, which the two companies began working with in 2024, Goodfire used Ember to build SAE-based probes that detect personally identifiable information before it is sent to downstream model providers; the resulting guardrail was deployed to Rakuten's agent platform, and Goodfire reports 15-500 times cost savings over LLM-as-a-judge setups with comparable performance ^[33].

On 9 October 2026 Goodfire and Baseten announced Project Beacon, which combines Goodfire's monitoring stack with Baseten's inference platform to detect unsafe agent behavior in real time for teams deploying open models. Goodfire said Baseten customers would be able to choose monitors for prompt injection, sensitive data exposure, cyber misuse, CBRNE misuse, and reward hacking ^[38].

Relationship with Anthropic

Goodfire's relationship with Anthropic has been important to its funding history. Its seed round included Menlo Ventures, investing through the Anthology Fund that Menlo co-leads with Anthropic ^[2][5]. In the Series A, Anthropic invested directly; Menlo described this as Anthropic's "first direct investment into another company" ^[2]. Anthropic is not listed among the Series B investors in Goodfire's announcement ^[3].

Dario Amodei framed the Series A investment as a bet that mechanistic interpretability can turn black-box networks into "understandable, steerable systems" ^[7]. Goodfire's own writing often cites Anthropic's work, including Amodei's essay "The Urgency of Interpretability" and Anthropic's Constitutional Classifiers++, which it describes as using probes as a cheap first stage in production monitoring ^[12][38]. In an August 2026 essay Balsam cited Anthropic's report that Constitutional Classifiers++ cut monitoring overhead from 23.7% to roughly 1%, and argued that deploying such guardrails across the industry is now mainly an engineering problem requiring changes to the inference stack ^[54].

Reception and significance

Goodfire has been one of the most prominent commercial efforts in mechanistic interpretability. MIT Technology Review, which named mechanistic interpretability one of its 10 Breakthrough Technologies of 2026, described Goodfire in April 2026 as "one of a small handful of companies, including industry leaders Anthropic, OpenAI, and Google DeepMind, pioneering" the technique ^[20][42]. The breakthrough-technologies entry itself credits Anthropic, Google DeepMind, Neuronpedia, and OpenAI and does not name Goodfire ^[42].

Outside commentators have questioned parts of Goodfire's positioning. Bereska told MIT Technology Review that calling Silico's approach engineering "makes it sound more principled than it is," while adding that such tools could help companies without in-house interpretability teams ^[20]. The reception of "Neural Networks Are Already Clean" split between researchers who praised it (one wrote that it "may be remembered as a major moment in scientific history") and skeptics such as Hilton, who wrote that it was "giving the public an unrealistic picture of the state of the field" ^[41]. Many of Goodfire's headline results, including the RLFR hallucination reduction, the cyber monitor figures, and the Alzheimer's biomarker findings, are self-reported and in some cases still awaiting peer review or experimental validation ^{[24][29][32][36]}.

Outlook

Goodfire's stated priorities as of October 2026 are foundational interpretability research and immediate safety applications. Its August 2026 grants announcement said the company had decided "to focus our research on two areas: interpretability foundations and safety applications," beginning with "identifying and preventing reward hacking, cybersecurity risks, and biosecurity risks directly from model internals" ^[21].

On 26 October 2026 McGrath and Mathilde Papillon published "Accelerating Interpretability: A Human Genome Project for AI," calling for "an ambitious, cross-industry effort to reverse-engineer a large language model" ^[43]. Goodfire committed $10 million to begin the project, with a first goal of unlocking "scaling laws for interpretability" by placing agents in an environment where they can run and verify experiments at scale, and an ultimate goal of fully reverse-engineering a frontier-scale model ^[43]. The essay says scaling to the most capable models "may require billions in shared costs" and that the output would be "a living encyclopedia, open to researchers everywhere" ^[43].

Timeline

Continue reading

• AnthropicAnthropic is an American artificial intelligence (AI) safety and research company founded in 2021 by Dario Amodei, Daniela Amodei, and other former OpenAI researchers, best known for the Claude family of large…Mentioned 20 times above
• LLaMALLaMA is the original name of a family of large language models developed by Meta AI.Mentioned 9 times above
• MMechanistic interpretabilityMechanistic interpretability (often abbreviated as mech interp or MI) is the field that reverse-engineers the internal computations of neural networks, particularly transformersMentioned 8 times above
• Google DeepMindGoogle DeepMind is an artificial intelligence research and model-development organization at Google.Mentioned 7 times above
• RReward hackingReward hacking (also called specification gaming) is a failure mode in artificial intelligence in which a system maximizes its given objective or reward signal through unintended shortcuts, exploits, or…Mentioned 6 times above
• Sparse autoencoderA sparse autoencoder (SAE) is a neural network that adds a sparsity penalty to an autoencoder's training loss so that only a small number of hidden units activate for any given input, producing a wideMentioned 5 times above

More from See also

• List of neolabs
• Neural Networks Are Already Clean
• Nick Cammarata
• Activation steering
• AI safety
• Stable Diffusion
• DeepSeek-R1
• Arc Institute
• Evo 2 (genomic model)

Topics:AI CompaniesAI SafetyResearch Organizations

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Goodfire. "Company." goodfire.com. Accessed 27 October 2026. goodfire.com/company
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Menlo Ventures. "Leading Goodfire's $50M Series A to Interpret How AI Models Think." April 2025. menlovc.com/...-a-to-interpret-how-ai-models-think
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13Goodfire. "Understanding, Learning From, and Designing AI: Our Series B." 5 February 2026. goodfire.com/...our-series-b
4. ^1 ^2Goodfire. "Careers." Accessed 27 October 2026. goodfire.com/careers
5. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Goodfire (press release, via Silicon UK). "Goodfire Raises $7M to Break Open the Black Box of Generative AI Models." August 2024. silicon.co.uk/...black-box-of-generative-ai-models
6. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Contrary Research. "Goodfire Business Breakdown & Founding Story." Updated 29 August 2025. research.contrary.com/...goodfire
7. ^1 ^2 ^3 ^4 ^5 ^6 ^7Goodfire. "Announcing Our $50M Series A to Advance AI Interpretability Research." 17 April 2025. goodfire.com/...announcing-our-50m-series-a
8. ^1 ^2 ^3Pillsbury Winthrop Shaw Pittman. "Pillsbury Represents Goodfire in $150 Million Series B Funding." 2026. pillsburylaw.com/...-goodfire-150-million-series-b
9. ^Work-Bench. "Goodfire Raises $150M Series B to Advance the Frontier of AI Interpretability." 2026. work-bench.com/...-frontier-of-ai-interpretability
10. ^Menlo Ventures. "Menlo Anthology Fund." menlovc.com/anthology-fund
11. ^Menlo Ventures. "Unveiling the First Cohort of Anthology Fund-Backed AI Startups." menlovc.com/...f-anthology-fund-backed-ai-startups
12. ^1 ^2 ^3 ^4 ^5 ^6 ^7McGrath, Thomas. "Intentionally Designing the Future of AI." Goodfire, 5 February 2026. goodfire.com/...intentional-design
13. ^1 ^2Ho, Eric. "We can and must solve alignment." Goodfire, 30 September 2026. goodfire.com/...we-can-and-must-solve-alignment
14. ^1 ^2 ^3 ^4 ^5Ho, Eric. "Feature Steering for Reliable and Expressive AI Engineering." Goodfire, 20 November 2024 (deprecation note added February 2026). goodfire.com/...able-and-expressive-ai-engineering
15. ^1 ^2 ^3 ^4Goodfire. "goodfire-sdk" (Ember API/SDK repository, archived). GitHub. github.com/...goodfire-sdk
16. ^1 ^2 ^3 ^4 ^5 ^6 ^7Balsam, D., McGrath, T., Gorton, L., Nguyen, N., Deng, M. "Announcing Open-Source SAEs for Llama 3.3 70B and Llama 3.1 8B." Goodfire, 10 January 2025. goodfire.com/...sae-open-source-announcement
17. ^1 ^2 ^3Hugging Face. "Goodfire" organization models. huggingface.co/Goodfire
18. ^1 ^2 ^3Cammarata, N., et al. "Painting With Concepts Using Diffusion Model Latents." Goodfire, 27 May 2025. goodfire.com/...painting-with-concepts
19. ^1 ^2Goodfire. "Our Approach to Safety at Goodfire." 23 December 2024. goodfire.com/...our-approach-to-safety
20. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Heaven, Will Douglas. "This startup's new mechanistic interpretability tool lets you debug LLMs." MIT Technology Review, 30 April 2026. technologyreview.com/...y-tool-lets-you-debug-llms
21. ^1 ^2 ^3 ^4 ^5 ^6 ^7Goodfire. "Announcing Goodfire Research Grants." 20 August 2026. goodfire.com/...announcing-goodfire-research-grants
22. ^1 ^2 ^3Goodfire. "How to build fast, efficient monitors for AI models using probes." 9 September 2026. goodfire.com/...probe-monitors-101
23. ^1 ^2 ^3 ^4 ^5 ^6 ^7Goodfire. "Research." Accessed 27 October 2026. goodfire.com/research
24. ^1 ^2 ^3 ^4 ^5Goodfire (Prasad et al.). "Features as Rewards: Using Interpretability to Reduce Hallucinations." 11 February 2026. goodfire.com/...rlfr
25. ^Prasad, A. V., Watts, C., Merullo, J., Gala, D., et al. "Features as Rewards: Scalable Supervision for Open-Ended Tasks via Interpretability." arXiv:2602.10067, 2026. arxiv.org/...2602.10067
26. ^1 ^2 ^3Goodfire (Hazra et al.). "Under the Hood of a Reasoning Model." 15 April 2025. goodfire.com/...under-the-hood-of-a-reasoning-model
27. ^1 ^2 ^3Goodfire (Gorton et al.). "Interpreting Evo 2: Arc Institute's Next-Generation Genomic Foundation Model." 20 February 2025. goodfire.com/...interpreting-evo-2
28. ^Brixi, G., et al. "Genome modelling and design across all domains of life with Evo 2." Nature 652, 1349-1361 (2026). doi.org/...s41586-026-10176-5
29. ^1 ^2 ^3Goodfire (Wang et al.). "Using Interpretability to Identify a Novel Class of Alzheimer's Biomarkers." 28 January 2026. goodfire.com/...etability-for-alzheimers-detection
30. ^1 ^2Goodfire. "Partnering with Radical AI to Advance Materials Science with Interpretability." 30 July 2025. goodfire.com/...radical-partnership-announcement
31. ^1 ^2 ^3Goodfire. "Goodfire Announces Collaboration to Advance Genomic Medicine with AI Interpretability." 9 September 2025. goodfire.com/...mayo-clinic-collaboration
32. ^1 ^2Goodfire (Pearce et al.). "Explaining 4.2 million genetic variants with state-of-the-art, interpretable predictions." 14 April 2026. goodfire.com/...evee-explaining-genetic-variants
33. ^1 ^2Goodfire. "How Rakuten secures reliable AI experiences for 44M+ monthly users" (customer story). goodfire.com/...rakuten
34. ^1 ^2 ^3Goodfire (Bergen et al.). "Models know when they're reward hacking, and we can catch them at scale." 17 September 2026. goodfire.com/...reward-hacking-activation-monitors
35. ^1 ^2Bergen, L., Bhalla, U., Lee, A., Widawsky, B., et al. "Monitoring and Discovering Reward Hacking with Internal Representations during LLM Evaluations." arXiv:2609.19101, 2026. arxiv.org/...2609.19101
36. ^1 ^2 ^3Goodfire (Lubana et al.). "Training and Deploying Production Cyber Monitors on Kimi K3." 8 October 2026. goodfire.com/...production-cyber-monitors-on-kimi-k3
37. ^Goodfire (Dooms et al.). "Better biosecurity monitors for AI agents via protein embeddings." 1 October 2026. goodfire.com/...better-biosecurity-monitors
38. ^1 ^2 ^3 ^4Goodfire. "Goodfire and Baseten partner to bring frontier safety to open models." 9 October 2026. goodfire.com/...goodfire-baseten-partnership
39. ^Anderson, M., Byun, M., Fross, T. "Interpretability Infrastructure at Frontier Scale: Harvesting Activations from a Trillion-Parameter Model." Goodfire, 25 February 2026. goodfire.com/...retability-infra-at-frontier-scale
40. ^1 ^2 ^3 ^4 ^5Cammarata, Nick. "Neural Networks Are Already Clean." Goodfire, 15 October 2026. already-clean.goodfire.com
41. ^1 ^2 ^3Heaven, Will Douglas. "'We were looking at it wrong': The explosive claim that's upended the AI world." MIT Technology Review, 19 October 2026. technologyreview.com/...thats-upended-the-ai-world
42. ^1 ^2Heaven, Will Douglas. "Mechanistic interpretability: 10 Breakthrough Technologies 2026." MIT Technology Review, 12 January 2026. technologyreview.com/...-breakthrough-technologies
43. ^1 ^2 ^3 ^4 ^5 ^6McGrath, T., Papillon, M. "Accelerating Interpretability: A Human Genome Project for AI." Goodfire, 26 October 2026. goodfire.com/...accelerating-interpretability
44. ^1 ^2Goodfire. "Announcing our SOC 2 Type II Compliance." 22 May 2026. goodfire.com/...soc-2-type-ii
45. ^1 ^2 ^3Goodfire. "Announcing Goodfire's Fellowship Program for Interpretability Research." 9 October 2025. goodfire.com/...fellowship-fall-25
46. ^1 ^2Bushnaq, L., Braun, D., Sharkey, L. "Stochastic Parameter Decomposition." arXiv:2506.20790, 2025. arxiv.org/...2506.20790
47. ^1 ^2Goodfire (Bushnaq et al.). "Interpreting Language Model Parameters." 5 May 2026. goodfire.com/...interpreting-lm-parameters
48. ^1 ^2Bergen, L., Bhalla, U., Baskaran, S., Loeffler, M., et al. "Anatomy of Post-Training: Using Interpretability to Characterize Data and Shape the Learning Signal." arXiv:2606.12360, 2026. arxiv.org/...2606.12360
49. ^Aranguri, S., McGrath, T. "Discovering Undesired Rare Behaviors via Model Diff Amplification." Goodfire, 21 August 2025. goodfire.com/...model-diff-amplification
50. ^Boppana, S., Ma, A., Loeffler, M., Sarfati, R., et al. "Reasoning Theater: Disentangling Model Beliefs from Chain-of-Thought." arXiv:2603.05488, 2026. arxiv.org/...2603.05488
51. ^Fel, T., Kowal, M., Jacobs, M., Hazra, D., et al. "Structuring Sparsity: Block-Sparse Featurizers Capture Visual Concept Manifolds." arXiv:2606.25234, 2026. arxiv.org/...2606.25234
52. ^1 ^2 ^3 ^4 ^5Goodfire. Homepage. Accessed 27 October 2026. goodfire.com
53. ^UC San Diego. "Leon Bergen" (faculty profile). profiles.ucsd.edu/leon.bergen
54. ^Balsam, Dan. "AI Safety Still Needs Great Engineers." Goodfire, 27 August 2026. goodfire.com/...safety-still-needs-great-engineers
55. ^1 ^2 ^3Goodfire. "How we identified a novel class of biomarkers for Alzheimer's detection" (Prima Mente customer story). goodfire.com/...prima-mente
56. ^Goodfire. "Life Sciences." Accessed 27 October 2026. goodfire.com/life-sciences

Show all 56 references