M&AAcquisitions & mergers
OpenAI Held Talks to Buy Interpretability Startup Goodfire for More Than $5 Billion
OpenAI explored buying Goodfire, an Anthropic-backed interpretability startup, for more than $5 billion, about four times its February valuation. Sam Altman was drawn to Nick Cammarata's contested essay arguing that the neurons in neural networks are far easier to interpret than researchers assumed. No deal resulted.
From the The Information article of November 16, 2026 and cited sources: A senior OpenAI executive reportedly floated an acquisition above $5 billion, or a partnership, in talks that Altman joined. [1] [2] [3] [4]
SOURCES[1] The Controversial AI Interpretability Research that Caught… theinformation.com[2] GuruFocus: OpenAI Explores Acquisition of Goodfire gurufocus.com[3] Newsquawk: OpenAI reportedly mulled $5bln Goodfire… newsquawk.com
AI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- A senior OpenAI executive reportedly floated an acquisition above $5 billion, or a partnership, in talks that Altman joined. [1] [2] [3] [4] [5]
- Original source
- The Information article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 16, 2026 article published
- Covered
- Nov 16 edition
- Evidence
- 11 sources across 10 sites
- 1The Controversial AI Interpretability Research that Caught…theinformation.com
- 2GuruFocus: OpenAI Explores Acquisition of Goodfiregurufocus.com
- 3Newsquawk: OpenAI reportedly mulled $5bln Goodfire…newsquawk.com
- 4Goodfire: Our Series Bgoodfire.com
- 5Pulse 2.0: Goodfire $150M Series Bpulse2.com
- 6The Decoder: Is AI interpretability close to a breakthrough?the-decoder.com
Reach: not measured yet
A senior OpenAI executive reportedly floated an acquisition above $5 billion, or a partnership, in talks that Altman joined. Goodfire had announced a $1.25 billion valuation with a $150 million Series B in February. It reportedly turned the deal down, expecting to be worth more by selling its tools to many AI developers. [1] [2] [3] [4] [5]
Altman's interest centered on Cammarata, a former OpenAI interpretability researcher. His essay argues that neurons responding to several unrelated concepts mostly reflect training failures and weak analysis tools, not superposition. This challenges a basic assumption behind sparse-autoencoder research. Neel Nanda and Eliezer Yudkowsky pushed back, while others called it a breakthrough. [1] [6] [7] [8] [11]
A follow-up technical report on Qwen3.6-35B-A3B claims that removing the weakest 92.2% of neuron activations raises loss by only 2.9%. It also says agent-written neuron descriptions separate strong firings from random text with 95% accuracy. These are Goodfire's own figures and have not been replicated. [9] [10]
The $5 billion figure comes from unnamed sources, while the $1.25 billion valuation was announced publicly.
Read the full assessment
If the research holds up, readable neurons could make model monitoring and debugging a strategic asset, which would help explain why a frontier lab is interested.
Executive brief
OpenAI reportedly discussed buying Goodfire, an interpretability startup partly funded by Anthropic, for more than $5 billion. That is over four times the $1.25 billion valuation Goodfire announced in February (GuruFocus; Newsquawk). No deal was reached. According to The Information, Sam Altman was especially interested in Nick Cammarata. Cammarata's October essay argues that the parts of AI models are much easier to read than the field believed. It has split researchers: some call it a breakthrough, others call it overhyped.
What changed and event timeline
Summer 2026: OpenAI holds talks with Goodfire
A senior OpenAI executive reportedly floated an acquisition above $5B, or a partnership. Altman joined some discussions. The figures come from unnamed sources (;).
Goodfire raises its Series B
The company announced a $150M round at a $1.25B valuation, led by B Capital, with Menlo, Lightspeed, Salesforce Ventures and Eric Schmidt also investing (;).
"Neural Networks Are Already Clean" is published
Cammarata's essay argues that "polysemantic" neurons, which respond to several unrelated concepts, mostly reflect training failures and weak analysis tools rather than a basic property of networks (;).
Researchers split
Roon (OpenAI) and Dean W. Ball praised the essay heavily. Neel Nanda (Google DeepMind) and Eliezer Yudkowsky pushed back ().
Companion technical report released
"Neurons as an Interpretable Basis" tests the essay's claims on Qwen3.6-35B-A3B (;).
The Information reports the talks
It links Altman's interest to Cammarata's research ().
Capabilities and access
- Models studied: InceptionV1, ConvNeXt, DINOv2, Qwen3-32B and Kimi K3 in the essay (essay).
- Access: both documents are free to read on the web. There is no released benchmark or replication package.
- Commercial product: Goodfire describes a "model design environment" for understanding model internals, training against them and monitoring models in production (Goodfire).
Read the full section
- Models studied: InceptionV1, ConvNeXt, DINOv2, Qwen3-32B and Kimi K3 in the essay (essay). Qwen3.6-35B-A3B in the report: about 5M neurons across 40 layers, mixture-of-experts architecture (report).
- Access: both documents are free to read on the web. There is no released benchmark or replication package.
- Commercial product: Goodfire describes a "model design environment" for understanding model internals, training against them and monitoring models in production (Goodfire).
Technical analysis for researchers and developers
- Core argument: the essay reinterprets the well-known "cat-car" neuron in InceptionV1 as a detector of V-shapes.
- Methods:
- Feature visualization, sparse autoencoders, steering and analysis of training checkpoints.
Read the full section
- Core argument: the essay reinterprets the well-known "cat-car" neuron in InceptionV1 as a detector of V-shapes. It proposes "recursive generalization" (a neuron branching into related concepts) as an alternative to superposition, the idea that models pack unrelated features into shared neurons (essay).
- Methods:
- Feature visualization, sparse autoencoders, steering and analysis of training checkpoints.
- Neuron labels come from researchers working by hand, combined with automated agents.
- The report also uses sparsification, where weak activations are removed, and attribution tools (report).
- What this means in practice:
- The headline percentages depend on where the threshold for "does one thing" is set.
- Explanations are tested mainly on strong activations, not weak ones.
- The results only matter at scale if per-neuron explanations generalize to other models and layers.
Claims and evidence
All figures below are reported by Goodfire and have not been independently replicated.
Read the full section
All figures below are reported by Goodfire and have not been independently replicated.
- Neuron function: about 90% of Qwen3-32B neurons have one main job, about 7% have two, and 2–4% are unexplained (The Decoder).
- Sparsity: removing the weakest 92.2% of neuron activations raises loss by only 2.9% (LessWrong).
- Explanation quality: descriptions written by agents distinguish strong firings from random text with 95% accuracy. Simulations based on them recover 73% of the damage caused by ablation. 94% of neurons can be pushed above their maximum observed activation (report).
- Earlier company claims: an approximately 50% reduction in hallucinations and a new class of Alzheimer's biomarkers (Goodfire).
Context and prior work
Cammarata worked on interpretability at OpenAI, and the essay rejects an explanation of neural network complexity he helped shape there (Trilogy AI). Superposition and sparse-autoencoder methods have been central to recent interpretability work, so the essay challenges a basic assumption of the field (The Decoder). Anthropic reportedly invested $1M in Goodfire, its first direct investment in another startup (GuruFocus).
Limitations, safety and contested findings
- Author's own caveats: the essay calls itself "a working hypothesis" and "highly speculative." It has no peer review and no held-out benchmark (Trilogy AI).
- Neel Nanda (Google DeepMind): superposition in the residual stream is still unsolved, and the essay's definition of a single-function neuron is loose (Zvi).
- Eliezer Yudkowsky: labeling individual neurons does not explain how a network solves problems (Zvi).
Read the full section
- Author's own caveats: the essay calls itself "a working hypothesis" and "highly speculative." It has no peer review and no held-out benchmark (Trilogy AI).
- Neel Nanda (Google DeepMind): superposition in the residual stream is still unsolved, and the essay's definition of a single-function neuron is loose (Zvi).
- Eliezer Yudkowsky: labeling individual neurons does not explain how a network solves problems (Zvi).
- The Colorado–Tornado neuron: a critic found that one of the essay's example neurons also fires strongly on "Tornado" code. The report answers that 98.8% of that neuron's causal effect comes from geography (LessWrong).
- Training risk: Zvi warns that training models directly on interpretability signals could teach them to hide their reasoning (Zvi).
Business and practitioner implications
- Valuation: a reported price above $5B, set against the $1.25B Goodfire announced itself, suggests frontier labs see strategic value in looking inside their models.
- Strategy: Goodfire reportedly turned down the deal because it expects to be worth more selling its tools to many AI developers (GuruFocus).
- For practitioners: neuron-level monitoring and sparsification look promising for debugging models.
Read the full section
- Valuation: a reported price above $5B, set against the $1.25B Goodfire announced itself, suggests frontier labs see strategic value in looking inside their models. Only the $1.25B figure is confirmed by the company and independent coverage.
- Strategy: Goodfire reportedly turned down the deal because it expects to be worth more selling its tools to many AI developers (GuruFocus).
- For practitioners: neuron-level monitoring and sparsification look promising for debugging models. Treat the headline percentages as hypotheses to test on your own models until outside groups reproduce them.
The source trail.
Sources (11)
The Controversial AI Interpretability Research that Caught Sam Altman’s Eye
Feed excerpt only.
theinformation.com