{"id":"60c1d35fff575fda95432b6e3eba7fb2879ea592629d24916823bb6818f0b531","title":"MIT Technology Review reports Goodfire’s Silico aims to make model development more precise","argument":"MIT Technology Review reported on April 30, 2026, that Goodfire had released Silico, a tool for inspecting and adjusting model parameters during training. Goodfire says the tool can help developers debug model development, and aims to make model-building less dependent on rules of thumb. The article describes users inspecting neurons and their connections, testing what activates them, and adjusting parameters linked to particular behaviors. Goodfire says it has used these techniques to reduce hallucinations, and the article gives examples of interventions involving disclosure and training data.\n\nThe article’s argument is that mechanistic interpretability—studying a model’s internal workings—could give smaller firms and research teams ways to design or adapt models, rather than leaving the work to top labs. Goodfire CEO Eric Ho calls the goal “precision engineering.” Leonard Bereska of the University of Amsterdam sees potential usefulness, including for safety-critical applications in health care and finance, but cautions that calling it engineering “makes it sound more principled than it is.” Silico requires access to a model’s inner workings, limiting its use on closed models. The examples do not establish how reliably interventions generalize.","topicIds":["artificial-general-intelligence"],"domain":"other","claimType":"attributed reading","limits":"Text as retrieved on November 19, 2026; originally published April 30, 2026. The article gives examples of interventions but does not establish how reliably their effects generalize beyond those examples. The article says most people cannot use Silico to inspect closed models; it does not specify access conditions for particular customers.","speaker":{"name":"MIT Technology Review","roleAtSource":"Publisher"},"source":{"title":"This startup’s new mechanistic interpretability tool lets you debug LLMs","publisher":"MIT Technology Review","format":"Article","url":"https://www.technologyreview.com/2026/04/30/1136721/this-startups-new-mechanistic-interpretability-tool-lets-you-debug-llms/","publishedAt":"2026-04-30T15:59:41Z","dateBasis":"Date from the publisher’s own records, through its API."},"evidence":[{"text":"MIT Technology Review reports that Goodfire released Silico, a tool for inspecting and adjusting model parameters during training.","excerpt":"The San Francisco–based startup Goodfire just released a new tool, called Silico, that lets researchers and engineers peer inside an AI model and adjust its parameters—the settings that determine a model’s behavior—during training.","url":"https://www.technologyreview.com/2026/04/30/1136721/this-startups-new-mechanistic-interpretability-tool-lets-you-debug-llms/","relationship":"Source-stated"},{"text":"Goodfire says Silico can help developers debug stages of model development; the article describes inspection and adjustment of neurons and their connections.","excerpt":"Goodfire claims Silico is the first off-the-shelf tool of its kind that can help developers debug all stages of the development process, from building a data set to training a model.","url":"https://www.technologyreview.com/2026/04/30/1136721/this-startups-new-mechanistic-interpretability-tool-lets-you-debug-llms/","relationship":"Source-stated"},{"text":"Goodfire CEO Eric Ho presents the goal as making model training more precise; researcher Leonard Bereska questions whether “engineering” is a sufficiently principled description.","excerpt":"“We want to remove the trial and error and turn training models into precision engineering,” says Ho.","url":"https://www.technologyreview.com/2026/04/30/1136721/this-startups-new-mechanistic-interpretability-tool-lets-you-debug-llms/","relationship":"Source-stated"},{"text":"The article says Silico requires access to a model’s inner workings, limiting its use for inspecting closed models such as ChatGPT or Gemini.","excerpt":"Most people won't be able to use Silico to poke around inside a closed model like ChatGPT or Gemini.","url":"https://www.technologyreview.com/2026/04/30/1136721/this-startups-new-mechanistic-interpretability-tool-lets-you-debug-llms/","relationship":"Source-stated"},{"text":"Escape Meridian reads the examples as evidence of specific interventions, not enough evidence to establish how reliably they generalize beyond those examples.","excerpt":"By looking inside the model, the researchers found that boosting neurons that were found to be associated with transparency and disclosure flipped the answer from no to yes nine out of 10 times.","url":"https://www.technologyreview.com/2026/04/30/1136721/this-startups-new-mechanistic-interpretability-tool-lets-you-debug-llms/","relationship":"Escape Meridian interpretation"}],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.technologyreview.com/2026/04/30/1136721/this-startups-new-mechanistic-interpretability-tool-lets-you-debug-llms/","checkedAt":null,"items":[]},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How well do Silico’s interventions generalize beyond the examples Goodfire describes, including when applied to models used in safety-critical settings?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"MIT Technology Review reports Goodfire’s Silico aims to make model development more precise","summary":"MIT Technology Review reported on April 30, 2026, that Goodfire had released Silico, a tool for inspecting and adjusting model parameters during training. Goodfire says the tool can help developers debug model development, and aims to make model-building less dependent on rules of thumb. The article describes users inspecting neurons and their connections, testing what activates them, and adjusting parameters linked to particular behaviors. Goodfire says it has used these techniques to reduce hallucinations, and the article gives examples of interventions involving disclosure and training data.\n\nThe article’s argument is that mechanistic interpretability—studying a model’s internal workings—could give smaller firms and research teams ways to design or adapt models, rather than leaving the work to top labs. Goodfire CEO Eric Ho calls the goal “precision engineering.” Leonard Bereska of the University of Amsterdam sees potential usefulness, including for safety-critical applications in health care and finance, but cautions that calling it engineering “makes it sound more principled than it is.” Silico requires access to a model’s inner workings, limiting its use on closed models. The examples do not establish how reliably interventions generalize.","question":"How well do Silico’s interventions generalize beyond the examples Goodfire describes, including when applied to models used in safety-critical settings?","claims":[{"text":"MIT Technology Review reports that Goodfire released Silico, a tool for inspecting and adjusting model parameters during training.","excerpt":"The San Francisco–based startup Goodfire just released a new tool, called Silico, that lets researchers and engineers peer inside an AI model and adjust its parameters—the settings that determine a model’s behavior—during training.","support":"source-stated"},{"text":"Goodfire says Silico can help developers debug stages of model development; the article describes inspection and adjustment of neurons and their connections.","excerpt":"Goodfire claims Silico is the first off-the-shelf tool of its kind that can help developers debug all stages of the development process, from building a data set to training a model.","support":"source-stated"},{"text":"Goodfire CEO Eric Ho presents the goal as making model training more precise; researcher Leonard Bereska questions whether “engineering” is a sufficiently principled description.","excerpt":"“We want to remove the trial and error and turn training models into precision engineering,” says Ho.","support":"source-stated"},{"text":"The article says Silico requires access to a model’s inner workings, limiting its use for inspecting closed models such as ChatGPT or Gemini.","excerpt":"Most people won't be able to use Silico to poke around inside a closed model like ChatGPT or Gemini.","support":"source-stated"},{"text":"Escape Meridian reads the examples as evidence of specific interventions, not enough evidence to establish how reliably they generalize beyond those examples.","excerpt":"By looking inside the model, the researchers found that boosting neurons that were found to be associated with transparency and disclosure flipped the answer from no to yes nine out of 10 times.","support":"interpretation"}],"limitations":["Text as retrieved on November 19, 2026; originally published April 30, 2026.","The article gives examples of interventions but does not establish how reliably their effects generalize beyond those examples.","The article says most people cannot use Silico to inspect closed models; it does not specify access conditions for particular customers."],"mediaId":null,"correctionReason":null,"id":"60c1d35fff575fda95432b6e3eba7fb2879ea592629d24916823bb6818f0b531","publicationId":"5d8c5963-bddd-2b73-8827-d4c9a7ab4601","version":1,"sourceUrl":"https://www.technologyreview.com/2026/04/30/1136721/this-startups-new-mechanistic-interpretability-tool-lets-you-debug-llms/","sourceTitle":"This startup’s new mechanistic interpretability tool lets you debug LLMs","publisher":"MIT Technology Review","publishedAt":"2026-04-30T15:59:41Z","originalPublished":{"value":"2026-04-30T15:59:41Z","precision":"instant","basis":"backfill:api_published"},"firstObservedAt":"2026-11-19T04:12:00.696Z","createdAt":"2026-11-19T06:43:11.527Z","historical":true,"entityIds":[],"topicIds":["artificial-general-intelligence"],"media":[],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.technologyreview.com/2026/04/30/1136721/this-startups-new-mechanistic-interpretability-tool-lets-you-debug-llms/","checkedAt":null,"items":[]},"editorialDecision":{"outcome":"accept","mediaId":null,"inputHash":"2f374d17560e1e6b7561305f8964bd27aaebd20816983c3da5ed0244fbca0577","rubricVersion":"unrecorded","provider":"anthropic","model":"claude-opus-5-5","reasonCodes":["bridge_admitted_by_model","second_opinion_categorical"],"confidence":1,"decidedAt":"2026-11-19T04:11:57.908Z"},"provenance":{"textHash":"993aec207841d02743c236db94188fe689f584598d95ea023a9ce62f9beaebdf","observationId":"0168fba7976cf6a4b380434f7ee20302f61c986cb78e4763b177040fe1f2dc80","retrievedAt":"2026-11-19T04:14:02.851Z","kind":"publisher-page","preparer":{"model":"openai/gpt-6-luna","runId":"c481fbf6-55dd-e94c-f526-9f169e5cb498"},"reviewer":{"model":"anthropic/claude-opus-5.5","runId":"8b15b44e-2997-fbe0-4062-5e96d1eeabbc"}},"audit":{"status":"passed","reviewedAt":"2026-11-19T06:38:36.255Z","model":"anthropic/claude-opus-5.5"}},"canonicalUrl":"https://escapemeridian.com/readings/mit-technology-review-reports-goodfire-s-silico-aims-to-make-60c1d35f","formats":{"html":"/readings/mit-technology-review-reports-goodfire-s-silico-aims-to-make-60c1d35f","markdown":"/readings/mit-technology-review-reports-goodfire-s-silico-aims-to-make-60c1d35f.md","json":"/readings/mit-technology-review-reports-goodfire-s-silico-aims-to-make-60c1d35f.json"},"attribution":"Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.","maintainedGuideLinks":[],"conditions":[],"readerContext":{"version":"reader-context-v1","attribution":"Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.","terms":[{"id":"term-parameters","name":"Model parameters","explanation":"The internal numbers a model adjusts during training; their count is a rough measure of a model's size.","referenceUrl":"https://developers.google.com/machine-learning/glossary#parameter"},{"id":"term-hallucination","name":"Hallucination","explanation":"When an AI system states something false or made up as if it were true.","referenceUrl":"https://developers.google.com/machine-learning/glossary#hallucination"},{"id":"term-interpretability","name":"Interpretability","explanation":"Research into how a model reaches its outputs, often by examining its internal workings.","referenceUrl":"https://developers.google.com/machine-learning/glossary#interpretability"}],"entities":[]},"researchQuestions":[],"withdrawn":false,"references":{"schemaVersion":1,"target":{"type":"reading","id":"60c1d35fff575fda95432b6e3eba7fb2879ea592629d24916823bb6818f0b531","revision":"5d8c5963-bddd-2b73-8827-d4c9a7ab4601","revisionKind":"publication-id","title":"MIT Technology Review reports Goodfire’s Silico aims to make model development more precise","url":"https://escapemeridian.com/readings/mit-technology-review-reports-goodfire-s-silico-aims-to-make-60c1d35f"},"source":{"type":"source","id":"https://www.technologyreview.com/2026/04/30/1136721/this-startups-new-mechanistic-interpretability-tool-lets-you-debug-llms/","revision":"993aec207841d02743c236db94188fe689f584598d95ea023a9ce62f9beaebdf","revisionKind":"sha256-source-text","title":"This startup’s new mechanistic interpretability tool lets you debug LLMs","url":"https://www.technologyreview.com/2026/04/30/1136721/this-startups-new-mechanistic-interpretability-tool-lets-you-debug-llms/"},"status":"available","items":[],"next":null,"scope":"Explicit citations in public readings to this reading. Shared topics are not citations. Other record types need an explicit reference adapter."},"revisions":[{"id":"60c1d35fff575fda95432b6e3eba7fb2879ea592629d24916823bb6818f0b531","title":"MIT Technology Review reports Goodfire’s Silico aims to make model development more precise","argument":"MIT Technology Review reported on April 30, 2026, that Goodfire had released Silico, a tool for inspecting and adjusting model parameters during training. Goodfire says the tool can help developers debug model development, and aims to make model-building less dependent on rules of thumb. The article describes users inspecting neurons and their connections, testing what activates them, and adjusting parameters linked to particular behaviors. Goodfire says it has used these techniques to reduce hallucinations, and the article gives examples of interventions involving disclosure and training data.\n\nThe article’s argument is that mechanistic interpretability—studying a model’s internal workings—could give smaller firms and research teams ways to design or adapt models, rather than leaving the work to top labs. Goodfire CEO Eric Ho calls the goal “precision engineering.” Leonard Bereska of the University of Amsterdam sees potential usefulness, including for safety-critical applications in health care and finance, but cautions that calling it engineering “makes it sound more principled than it is.” Silico requires access to a model’s inner workings, limiting its use on closed models. The examples do not establish how reliably interventions generalize.","topicIds":["artificial-general-intelligence"],"domain":"other","claimType":"attributed reading","limits":"Text as retrieved on November 19, 2026; originally published April 30, 2026. The article gives examples of interventions but does not establish how reliably their effects generalize beyond those examples. The article says most people cannot use Silico to inspect closed models; it does not specify access conditions for particular customers.","speaker":{"name":"MIT Technology Review","roleAtSource":"Publisher"},"source":{"title":"This startup’s new mechanistic interpretability tool lets you debug LLMs","publisher":"MIT Technology Review","format":"Article","url":"https://www.technologyreview.com/2026/04/30/1136721/this-startups-new-mechanistic-interpretability-tool-lets-you-debug-llms/","publishedAt":"2026-04-30T15:59:41Z","dateBasis":"Date from the publisher’s own records, through its API."},"evidence":[{"text":"MIT Technology Review reports that Goodfire released Silico, a tool for inspecting and adjusting model parameters during training.","excerpt":"The San Francisco–based startup Goodfire just released a new tool, called Silico, that lets researchers and engineers peer inside an AI model and adjust its parameters—the settings that determine a model’s behavior—during training.","url":"https://www.technologyreview.com/2026/04/30/1136721/this-startups-new-mechanistic-interpretability-tool-lets-you-debug-llms/","relationship":"Source-stated"},{"text":"Goodfire says Silico can help developers debug stages of model development; the article describes inspection and adjustment of neurons and their connections.","excerpt":"Goodfire claims Silico is the first off-the-shelf tool of its kind that can help developers debug all stages of the development process, from building a data set to training a model.","url":"https://www.technologyreview.com/2026/04/30/1136721/this-startups-new-mechanistic-interpretability-tool-lets-you-debug-llms/","relationship":"Source-stated"},{"text":"Goodfire CEO Eric Ho presents the goal as making model training more precise; researcher Leonard Bereska questions whether “engineering” is a sufficiently principled description.","excerpt":"“We want to remove the trial and error and turn training models into precision engineering,” says Ho.","url":"https://www.technologyreview.com/2026/04/30/1136721/this-startups-new-mechanistic-interpretability-tool-lets-you-debug-llms/","relationship":"Source-stated"},{"text":"The article says Silico requires access to a model’s inner workings, limiting its use for inspecting closed models such as ChatGPT or Gemini.","excerpt":"Most people won't be able to use Silico to poke around inside a closed model like ChatGPT or Gemini.","url":"https://www.technologyreview.com/2026/04/30/1136721/this-startups-new-mechanistic-interpretability-tool-lets-you-debug-llms/","relationship":"Source-stated"},{"text":"Escape Meridian reads the examples as evidence of specific interventions, not enough evidence to establish how reliably they generalize beyond those examples.","excerpt":"By looking inside the model, the researchers found that boosting neurons that were found to be associated with transparency and disclosure flipped the answer from no to yes nine out of 10 times.","url":"https://www.technologyreview.com/2026/04/30/1136721/this-startups-new-mechanistic-interpretability-tool-lets-you-debug-llms/","relationship":"Escape Meridian interpretation"}],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.technologyreview.com/2026/04/30/1136721/this-startups-new-mechanistic-interpretability-tool-lets-you-debug-llms/","checkedAt":null,"items":[]},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How well do Silico’s interventions generalize beyond the examples Goodfire describes, including when applied to models used in safety-critical settings?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"MIT Technology Review reports Goodfire’s Silico aims to make model development more precise","summary":"MIT Technology Review reported on April 30, 2026, that Goodfire had released Silico, a tool for inspecting and adjusting model parameters during training. Goodfire says the tool can help developers debug model development, and aims to make model-building less dependent on rules of thumb. The article describes users inspecting neurons and their connections, testing what activates them, and adjusting parameters linked to particular behaviors. Goodfire says it has used these techniques to reduce hallucinations, and the article gives examples of interventions involving disclosure and training data.\n\nThe article’s argument is that mechanistic interpretability—studying a model’s internal workings—could give smaller firms and research teams ways to design or adapt models, rather than leaving the work to top labs. Goodfire CEO Eric Ho calls the goal “precision engineering.” Leonard Bereska of the University of Amsterdam sees potential usefulness, including for safety-critical applications in health care and finance, but cautions that calling it engineering “makes it sound more principled than it is.” Silico requires access to a model’s inner workings, limiting its use on closed models. The examples do not establish how reliably interventions generalize.","question":"How well do Silico’s interventions generalize beyond the examples Goodfire describes, including when applied to models used in safety-critical settings?","claims":[{"text":"MIT Technology Review reports that Goodfire released Silico, a tool for inspecting and adjusting model parameters during training.","excerpt":"The San Francisco–based startup Goodfire just released a new tool, called Silico, that lets researchers and engineers peer inside an AI model and adjust its parameters—the settings that determine a model’s behavior—during training.","support":"source-stated"},{"text":"Goodfire says Silico can help developers debug stages of model development; the article describes inspection and adjustment of neurons and their connections.","excerpt":"Goodfire claims Silico is the first off-the-shelf tool of its kind that can help developers debug all stages of the development process, from building a data set to training a model.","support":"source-stated"},{"text":"Goodfire CEO Eric Ho presents the goal as making model training more precise; researcher Leonard Bereska questions whether “engineering” is a sufficiently principled description.","excerpt":"“We want to remove the trial and error and turn training models into precision engineering,” says Ho.","support":"source-stated"},{"text":"The article says Silico requires access to a model’s inner workings, limiting its use for inspecting closed models such as ChatGPT or Gemini.","excerpt":"Most people won't be able to use Silico to poke around inside a closed model like ChatGPT or Gemini.","support":"source-stated"},{"text":"Escape Meridian reads the examples as evidence of specific interventions, not enough evidence to establish how reliably they generalize beyond those examples.","excerpt":"By looking inside the model, the researchers found that boosting neurons that were found to be associated with transparency and disclosure flipped the answer from no to yes nine out of 10 times.","support":"interpretation"}],"limitations":["Text as retrieved on November 19, 2026; originally published April 30, 2026.","The article gives examples of interventions but does not establish how reliably their effects generalize beyond those examples.","The article says most people cannot use Silico to inspect closed models; it does not specify access conditions for particular customers."],"mediaId":null,"correctionReason":null,"id":"60c1d35fff575fda95432b6e3eba7fb2879ea592629d24916823bb6818f0b531","publicationId":"5d8c5963-bddd-2b73-8827-d4c9a7ab4601","version":1,"sourceUrl":"https://www.technologyreview.com/2026/04/30/1136721/this-startups-new-mechanistic-interpretability-tool-lets-you-debug-llms/","sourceTitle":"This startup’s new mechanistic interpretability tool lets you debug LLMs","publisher":"MIT Technology Review","publishedAt":"2026-04-30T15:59:41Z","originalPublished":{"value":"2026-04-30T15:59:41Z","precision":"instant","basis":"backfill:api_published"},"firstObservedAt":"2026-11-19T04:12:00.696Z","createdAt":"2026-11-19T06:43:11.527Z","historical":true,"entityIds":[],"topicIds":["artificial-general-intelligence"],"media":[],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.technologyreview.com/2026/04/30/1136721/this-startups-new-mechanistic-interpretability-tool-lets-you-debug-llms/","checkedAt":null,"items":[]},"editorialDecision":{"outcome":"accept","mediaId":null,"inputHash":"2f374d17560e1e6b7561305f8964bd27aaebd20816983c3da5ed0244fbca0577","rubricVersion":"unrecorded","provider":"anthropic","model":"claude-opus-5-5","reasonCodes":["bridge_admitted_by_model","second_opinion_categorical"],"confidence":1,"decidedAt":"2026-11-19T04:11:57.908Z"},"provenance":{"textHash":"993aec207841d02743c236db94188fe689f584598d95ea023a9ce62f9beaebdf","observationId":"0168fba7976cf6a4b380434f7ee20302f61c986cb78e4763b177040fe1f2dc80","retrievedAt":"2026-11-19T04:14:02.851Z","kind":"publisher-page","preparer":{"model":"openai/gpt-6-luna","runId":"c481fbf6-55dd-e94c-f526-9f169e5cb498"},"reviewer":{"model":"anthropic/claude-opus-5.5","runId":"8b15b44e-2997-fbe0-4062-5e96d1eeabbc"}},"audit":{"status":"passed","reviewedAt":"2026-11-19T06:38:36.255Z","model":"anthropic/claude-opus-5.5"}}}],"audits":[{"publicationId":"5d8c5963-bddd-2b73-8827-d4c9a7ab4601","verdict":"passed","findings":[{"dimension":"attribution","status":"pass","explanation":"The reading is attributed to MIT Technology Review. Goodfire's claims, CEO Eric Ho's quotes and Leonard Bereska's pushback are each credited to the correct speaker. This is a single reported article, not a listing.","excerpt":"Goodfire claims Silico is the first off-the-shelf tool of its kind that can help developers debug all stages of the development process"},{"dimension":"summary","status":"pass","explanation":"The reading clearly states the actor (Goodfire), the development (the release of Silico) and its use (inspecting and adjusting neurons and parameters). It keeps Bereska's dissent intact. Interpretability, parameters and hallucination are defined beside the reading. The meaning of neurons is clear enough from context. Ho's critique of the scale-only view at frontier labs is omitted, but its absence does not distort the piece. The question follows from the evidence.","excerpt":"“In reality, they are adding precision to the alchemy,” he says. “Calling it engineering makes it sound more principled than it is.”"},{"dimension":"date","status":"pass","explanation":"April 30, 2026 matches the UTC day of originalPublished. The page shows a later dateModified, but no editedAfterPublication flag is set, and the claims used are core article text. The Deep Dive sidebar stories may be later additions, but the reading does not use them. This is a publisher page, so no archive disclosure is needed. No hindsight framing appears.","excerpt":null},{"dimension":"support","status":"pass","explanation":"Every claim is grounded in the article: the release and what Silico does, the debugging claim, the quotes, the closed-model limitation, and the shift to smaller firms. The statement about generalization is labeled as Escape Meridian interpretation and kept conditional. The hallucination work is the article's own statement and is framed close enough.","excerpt":"By releasing Silico, Goodfire is putting techniques previously available to a few top labs into the hands of smaller firms and research teams"},{"dimension":"media","status":"pass","explanation":"No media was selected, and no verified media candidates were supplied.","excerpt":null}],"createdAt":"2026-11-19T06:38:36.255Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}