Here\u2019s a demo of HFlow in action: https://www.youtube.com/watch?v=xni0GwV-xAw
Robotics data pipelines often begin as scripts: one transcodes video, another checks timestamps, another adds labels, and another copies selected recordings into a training set. This works until the corpus grows. Then it becomes difficult to know which code ran, why an episode was excluded, or whether a dataset can be reproduced. The first pain is usually quality control because frozen cameras, missing topics, timestamp drift, and duplicate recordings can quietly enter training data.
Brandon first encountered this while training embodied AI models for two-arm industrial cleaning robots. Kingston had run into related problems while building high-throughput infrastructure at Jane Street. Later, while speaking with robotics data providers, we kept seeing teams rebuild similar processing and quality-control infrastructure. We learnt that processing robotics data is itself one of the bottlenecks to improving robotics models.
An HFlow pipeline consists of transformations, checks, labels, and enrichments. The SDK exposes them as plain Python functions that receive an episode and return measurements, artifacts, or transformed data. During development, the functions can run in-process. For scheduled corpus processing, HFlow packages the same registered steps as Airflow 3 DAGs, where teams can inspect task status, logs, retries, and reruns.
HFlow currently accepts one MCAP file per episode. MCAP (https://mcap.dev/) is an open container format by Foxglove for timestamped multimodal recordings, similar in purpose to a ROS bag. It lets video, robot state, actions, and other sensor streams remain synchronized in one file. We use it because HFlow needs to process these streams together, and because the resulting recordings remain compatible with Foxglove and Rerun. HFlow writes a canonical MCAP with in-band H.264 video, grouped camera and state chunks, and provenance describing how the output was produced. Each step has an explicit behavior version, and catalog records connect its measurements and artifacts to the source episode and pipeline run.
Quality checks store reusable evidence rather than imposing one universal definition of good data. Some failures, including black frames, frozen video, missing topics, timestamp drift, and impossible joint movements, can be measured deterministically without training a model. Others might be detected using VLMs and other models like MediaPipe Hands. But their meaning depends on the task. A smooth trajectory might indicate a successful demonstration in one setting and a stalled robot in another.
HFlow writes measurements, metadata, version stamps, and artifact locations to an append-only Parquet catalog. Teams query it with DuckDB SQL and produce a version-pinned manifest without opening the recordings again. Critical checks can quarantine an episode, but HFlow does not delete data. This separates the evidence from the policy used to assemble a particular dataset.
We did not want to replace the tools robotics teams already use. HFlow connects MCAP for synchronized recordings, Airflow for scheduled execution, Parquet for catalog data, and DuckDB for curation. Compared with a general workflow orchestrator, it adds contracts for robotics episodes, processing provenance, quality evidence, quarantine, and dataset manifests. Compared with a training dataset format, it operates earlier and stops at curated episodes plus a manifest.
Here are three examples of teams that would use HFlow:\n1. A data vendor or marketplace collecting egocentric recordings. They could use HFlow to detect black or frozen video, duplicate recordings, hand-object interaction, and other quality metrics before delivering the data, while retaining evidence of which checks ran on every episode.\n2. A robotics team collecting teleoperated demonstrations for its own models. They could use HFlow to standardize recordings, add labels and enrichments, and produce a reproducible training manifest.\n3. A team operating robots in the field. It could process incoming logs, quarantine incomplete or corrupted episodes, and query the catalog for particular robot versions, environments, or failure conditions.
The project is pre-v1, but the core lifecycle works end to end. You can try it without an account, Docker, or robot hardware by cloning the repository and following the quickstart.
HFlow is free under the Apache-2.0 license. The open source deployment is currently a single-tenant workspace, and we have not built the hosted, multi-tenant control plane yet. We are considering making money through managed workspaces and enterprise support for teams that do not want to operate the runtime themselves.
Because this processing layer is software and data, people can contribute without owning a robot. We would especially like feedback from people who have built pipelines for robotics, video, or other sensor-heavy systems. We are curious where our data model is wrong, which integrations are missing, and what would fail first on your workloads."},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["hebbia"],"value":"Launch HN: Hebbian Robotics (YC S26) \u2013 Build scalable robotics data pipelines"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["hebbia"],"value":"https://github.com/Hebbian-Robotics/hflow"}},"_tags":["story","author_kstonekuan","story_49510632","launch_hn"],"author":"kstonekuan","children":[49511266,49511285,49511471,49512337,49512664],"created_at":"2026-08-31T15:02:41Z","created_at_i":1788188561,"num_comments":12,"objectID":"49510632","points":46,"story_id":49510632,"story_text":"Hi HN, we\u2019re Brandon and Kingston, the founders of Hebbian Robotics. We built HFlow (https://github.com/Hebbian-Robotics/hflow), an SDK that turns multimodal recordings from robots and human operators into standardized, quality-checked episodes and queryable dataset manifests. A recording can contain synchronized video, joint states, actions, timestamps, and metadata, and HFlow processes those streams together.
Here\u2019s a demo of HFlow in action: https://www.youtube.com/watch?v=xni0GwV-xAw
Robotics data pipelines often begin as scripts: one transcodes video, another checks timestamps, another adds labels, and another copies selected recordings into a training set. This works until the corpus grows. Then it becomes difficult to know which code ran, why an episode was excluded, or whether a dataset can be reproduced. The first pain is usually quality control because frozen cameras, missing topics, timestamp drift, and duplicate recordings can quietly enter training data.
Brandon first encountered this while training embodied AI models for two-arm industrial cleaning robots. Kingston had run into related problems while building high-throughput infrastructure at Jane Street. Later, while speaking with robotics data providers, we kept seeing teams rebuild similar processing and quality-control infrastructure. We learnt that processing robotics data is itself one of the bottlenecks to improving robotics models.
An HFlow pipeline consists of transformations, checks, labels, and enrichments. The SDK exposes them as plain Python functions that receive an episode and return measurements, artifacts, or transformed data. During development, the functions can run in-process. For scheduled corpus processing, HFlow packages the same registered steps as Airflow 3 DAGs, where teams can inspect task status, logs, retries, and reruns.
HFlow currently accepts one MCAP file per episode. MCAP (https://mcap.dev/) is an open container format by Foxglove for timestamped multimodal recordings, similar in purpose to a ROS bag. It lets video, robot state, actions, and other sensor streams remain synchronized in one file. We use it because HFlow needs to process these streams together, and because the resulting recordings remain compatible with Foxglove and Rerun. HFlow writes a canonical MCAP with in-band H.264 video, grouped camera and state chunks, and provenance describing how the output was produced. Each step has an explicit behavior version, and catalog records connect its measurements and artifacts to the source episode and pipeline run.
Quality checks store reusable evidence rather than imposing one universal definition of good data. Some failures, including black frames, frozen video, missing topics, timestamp drift, and impossible joint movements, can be measured deterministically without training a model. Others might be detected using VLMs and other models like MediaPipe Hands. But their meaning depends on the task. A smooth trajectory might indicate a successful demonstration in one setting and a stalled robot in another.
HFlow writes measurements, metadata, version stamps, and artifact locations to an append-only Parquet catalog. Teams query it with DuckDB SQL and produce a version-pinned manifest without opening the recordings again. Critical checks can quarantine an episode, but HFlow does not delete data. This separates the evidence from the policy used to assemble a particular dataset.
We did not want to replace the tools robotics teams already use. HFlow connects MCAP for synchronized recordings, Airflow for scheduled execution, Parquet for catalog data, and DuckDB for curation. Compared with a general workflow orchestrator, it adds contracts for robotics episodes, processing provenance, quality evidence, quarantine, and dataset manifests. Compared with a training dataset format, it operates earlier and stops at curated episodes plus a manifest.
Here are three examples of teams that would use HFlow:\n1. A data vendor or marketplace collecting egocentric recordings. They could use HFlow to detect black or frozen video, duplicate recordings, hand-object interaction, and other quality metrics before delivering the data, while retaining evidence of which checks ran on every episode.\n2. A robotics team collecting teleoperated demonstrations for its own models. They could use HFlow to standardize recordings, add labels and enrichments, and produce a reproducible training manifest.\n3. A team operating robots in the field. It could process incoming logs, quarantine incomplete or corrupted episodes, and query the catalog for particular robot versions, environments, or failure conditions.
The project is pre-v1, but the core lifecycle works end to end. You can try it without an account, Docker, or robot hardware by cloning the repository and following the quickstart.
HFlow is free under the Apache-2.0 license. The open source deployment is currently a single-tenant workspace, and we have not built the hosted, multi-tenant control plane yet. We are considering making money through managed workspaces and enterprise support for teams that do not want to operate the runtime themselves.
Because this processing layer is software and data, people can contribute without owning a robot. We would especially like feedback from people who have built pipelines for robotics, video, or other sensor-heavy systems. We are curious where our data model is wrong, which integrations are missing, and what would fail first on your workloads.","title":"Launch HN: Hebbian Robotics (YC S26) \u2013 Build scalable robotics data pipelines","updated_at":"2026-09-07T02:39:12Z","url":"https://github.com/Hebbian-Robotics/hflow"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"srevanur"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["hebbia"],"value":"GPT-3, Stable Diffusion, DALL-E, and most recently, Chat-GPT have thrust AI\u2019s ability to create into the mainstream.
But creations alone are often a long way from being helpful \u2013Chat-GPT\u2019s incorrect answers, and the immediate takedown of Meta\u2019s Galactica share a glaring flaw:
Generative models can write responses, but they\u2019re never reading sources or \u201cciting\u201d their work.
At Hebbia, we\u2019re building semantic search \u201creaders\u201d that feed generative \u201cwriters\u201d to address these shortcomings of generative models alone:
1. More accurate: Priming models with relevant primary sources
2. More trustworthy: Cite sources behind every generation
3. More easily updated: Updating an index can happen almost instantaneously with no need to retrain a billion (or two trillion!) parameter model ad hoc.
Read more: https://medium.com/@gsiv/semantic-search-chatgpt-729d71263851
Build the future of knowledge engines with us: http://careers.hebbia.ai/"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["hebbia"],"value":"ChatGPT writes. Hebbia reads. Why they\u2019re a great pair."}},"_tags":["story","author_srevanur","story_33994795","ask_hn"],"author":"srevanur","children":[33996945],"created_at":"2022-12-15T03:30:28Z","created_at_i":1671075028,"num_comments":1,"objectID":"33994795","points":13,"story_id":33994795,"story_text":"GPT-3, Stable Diffusion, DALL-E, and most recently, Chat-GPT have thrust AI\u2019s ability to create into the mainstream.
But creations alone are often a long way from being helpful \u2013Chat-GPT\u2019s incorrect answers, and the immediate takedown of Meta\u2019s Galactica share a glaring flaw:
Generative models can write responses, but they\u2019re never reading sources or \u201cciting\u201d their work.
At Hebbia, we\u2019re building semantic search \u201creaders\u201d that feed generative \u201cwriters\u201d to address these shortcomings of generative models alone:
1. More accurate: Priming models with relevant primary sources
2. More trustworthy: Cite sources behind every generation
3. More easily updated: Updating an index can happen almost instantaneously with no need to retrain a billion (or two trillion!) parameter model ad hoc.
Read more: https://medium.com/@gsiv/semantic-search-chatgpt-729d71263851
Build the future of knowledge engines with us: http://careers.hebbia.ai/","title":"ChatGPT writes. Hebbia reads. Why they\u2019re a great pair.","updated_at":"2024-09-20T12:54:44Z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"jterwill"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["hebbia"],"value":"Show HN: Interactive computational neuroscience: synapses and Hebbian learning"},"url":{"matchLevel":"none","matchedWords":[],"value":"http://jackterwilliger.com/biological-neural-network-synapses/"}},"_tags":["story","author_jterwill","story_17834721","show_hn"],"author":"jterwill","created_at":"2018-08-24T13:15:16Z","created_at_i":1535116516,"num_comments":0,"objectID":"17834721","points":9,"story_id":17834721,"title":"Show HN: Interactive computational neuroscience: synapses and Hebbian learning","updated_at":"2024-09-20T02:56:09Z","url":"http://jackterwilliger.com/biological-neural-network-synapses/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"gsiv"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["hebbia"],"value":"Goodbye RAG \u2013 How Hebbia Solved Information Retrieval for LLMs"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["hebbia"],"value":"https://www.hebbia.com/blog/goodbye-rag-how-hebbia-solved-information-retrieval-for-llms"}},"_tags":["story","author_gsiv","story_43107258"],"author":"gsiv","created_at":"2025-02-19T20:32:45Z","created_at_i":1739997165,"num_comments":0,"objectID":"43107258","points":8,"story_id":43107258,"title":"Goodbye RAG \u2013 How Hebbia Solved Information Retrieval for LLMs","updated_at":"2025-03-18T07:03:47Z","url":"https://www.hebbia.com/blog/goodbye-rag-how-hebbia-solved-information-retrieval-for-llms"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"jterwill"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["hebbia"],"value":" Interactive computational neuroscience \u2013 part 2, synapses & hebbian learning"},"url":{"matchLevel":"none","matchedWords":[],"value":"http://jackterwilliger.com/biological-neural-network-synapses/"}},"_tags":["story","author_jterwill","story_17810174"],"author":"jterwill","created_at":"2018-08-21T15:00:59Z","created_at_i":1534863659,"num_comments":0,"objectID":"17810174","points":8,"story_id":17810174,"title":" Interactive computational neuroscience \u2013 part 2, synapses & hebbian learning","updated_at":"2024-09-20T02:53:32Z","url":"http://jackterwilliger.com/biological-neural-network-synapses/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"tgragnato"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["hebbia"],"value":"Anti-Hebbian and Hebbian (AHaH) Plasticity: Neuromemristive Processors"},"url":{"matchLevel":"none","matchedWords":[],"value":"http://knowm.org/ahah-computing/"}},"_tags":["story","author_tgragnato","story_14009759"],"author":"tgragnato","created_at":"2017-04-01T01:09:54Z","created_at_i":1491008994,"num_comments":0,"objectID":"14009759","points":4,"story_id":14009759,"title":"Anti-Hebbian and Hebbian (AHaH) Plasticity: Neuromemristive Processors","updated_at":"2024-09-20T00:33:02Z","url":"http://knowm.org/ahah-computing/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"PaulHoule"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["hebbia"],"value":"Unsupervised Hebbian Learning from an Artificial Intelligence Perspectives"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://www.mdpi.com/2504-4990/7/4/143"}},"_tags":["story","author_PaulHoule","story_46040084"],"author":"PaulHoule","created_at":"2025-11-24T22:20:53Z","created_at_i":1764022853,"num_comments":0,"objectID":"46040084","points":3,"story_id":46040084,"title":"Unsupervised Hebbian Learning from an Artificial Intelligence Perspectives","updated_at":"2026-03-05T23:06:57Z","url":"https://www.mdpi.com/2504-4990/7/4/143"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"mjbonanno"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["hebbia"],"value":"Hi HN,
After building several AI agent systems, I kept running into the same frustration: memory layers that are either static vector stores or fragile prompt hacks. Retrieval is opaque, forgetting happens at the wrong time, and associations don\u2019t form naturally.
So I threw away the two production memory systems I had and built something different. MuninnDB is a purpose-built cognitive memory database where memories (called engrams) are first-class citizens that:
- Strengthen with repeated co-activation (Hebbian learning)\n- Decay over time using a verbatim ACT-R formula\n- Automatically form bidirectional associations\n- Track their own Bayesian confidence\n- Return a full mathematical \u201cWhy\u201d explanation on every retrieval
Everything runs as a single static Go binary (embedded Pebble LSM storage + HNSW + BM25). No external services, no Redis/Postgres/Pinecone, and no LLM in the hot path. One command (muninn init) auto-configures it with Cursor, Claude Desktop, VS Code, and any other MCP-compatible tool.
The core call is dead simple: Activate(context) returns ranked results + explainable scoring. Background workers handle learning and decay on every read.
GitHub: https://github.com/scrypster/muninndb\nWebsite + docs + install (one-liner): https://muninndb.com\nQuick 13-minute demo video: https://www.youtube.com/watch?v=b29wl0ehrQI
It\u2019s very early (alpha, ~10 days old), but already functional and I\u2019m using it daily. Would love honest feedback or questions from anyone working on agent memory, long-term RAG, or cognitive architectures.
Thanks!"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["hebbia"],"value":"Show HN: MuninnDB \u2013 ACT-R decay and Hebbian memory for AI agents"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/scrypster/muninndb"}},"_tags":["story","author_mjbonanno","story_47236100","show_hn"],"author":"mjbonanno","children":[47236108,47236134,47236252,47239863],"created_at":"2026-03-03T17:53:00Z","created_at_i":1772560380,"num_comments":7,"objectID":"47236100","points":2,"story_id":47236100,"story_text":"Hi HN,
After building several AI agent systems, I kept running into the same frustration: memory layers that are either static vector stores or fragile prompt hacks. Retrieval is opaque, forgetting happens at the wrong time, and associations don\u2019t form naturally.
So I threw away the two production memory systems I had and built something different. MuninnDB is a purpose-built cognitive memory database where memories (called engrams) are first-class citizens that:
- Strengthen with repeated co-activation (Hebbian learning)\n- Decay over time using a verbatim ACT-R formula\n- Automatically form bidirectional associations\n- Track their own Bayesian confidence\n- Return a full mathematical \u201cWhy\u201d explanation on every retrieval
Everything runs as a single static Go binary (embedded Pebble LSM storage + HNSW + BM25). No external services, no Redis/Postgres/Pinecone, and no LLM in the hot path. One command (muninn init) auto-configures it with Cursor, Claude Desktop, VS Code, and any other MCP-compatible tool.
The core call is dead simple: Activate(context) returns ranked results + explainable scoring. Background workers handle learning and decay on every read.
GitHub: https://github.com/scrypster/muninndb\nWebsite + docs + install (one-liner): https://muninndb.com\nQuick 13-minute demo video: https://www.youtube.com/watch?v=b29wl0ehrQI
It\u2019s very early (alpha, ~10 days old), but already functional and I\u2019m using it daily. Would love honest feedback or questions from anyone working on agent memory, long-term RAG, or cognitive architectures.
Thanks!","title":"Show HN: MuninnDB \u2013 ACT-R decay and Hebbian memory for AI agents","updated_at":"2026-03-27T11:46:53Z","url":"https://github.com/scrypster/muninndb"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"ELAmrani"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["hebbia"],"value":"Hi HN,
I have been tinkering on a project for quite some time now.\nMost AI finance tools begin after the data is ready. Financial Diligence and audits don\u2019t.
I built SecondState to start one layer earlier.
The interesting pieces:
- Agentic data transformation: It has kind of cursor for \u201cdata engineers\u201d : you declare what you want and the agent builds the DAG and materialize it with a human-in-the-loop approval gate.
- Just-in-time Analytics: once you have some clean data, you can ask whatever question and it builds a dashboard on demand, helping users narrow the aperture on what\u2019s valuable to look at.
The problem originates from my previous experience at a BIG 4 :
Every deal starts with a data room: General ledgers. Accounting records. Contracts. Financial statements & Endless spreadsheets.
And before anyone can understand operate on top, a loot of offshore teams in poor countries have to reconstruct it by hand and inject it into a dashboard.
Does this kill Hebbia? No. But I think the winning platform for high-stakes financial work will need to do more than RAG and knowledge graphs."},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["hebbia"],"value":"Show HN: SecondState \u2013 what if Hebbia started with raw ERP data?"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://trysecondstate.com/"}},"_tags":["story","author_ELAmrani","story_49251803","show_hn"],"author":"ELAmrani","children":[49258587],"created_at":"2026-08-11T00:39:08Z","created_at_i":1786408748,"num_comments":0,"objectID":"49251803","points":2,"story_id":49251803,"story_text":"Hi HN,
I have been tinkering on a project for quite some time now.\nMost AI finance tools begin after the data is ready. Financial Diligence and audits don\u2019t.
I built SecondState to start one layer earlier.
The interesting pieces:
- Agentic data transformation: It has kind of cursor for \u201cdata engineers\u201d : you declare what you want and the agent builds the DAG and materialize it with a human-in-the-loop approval gate.
- Just-in-time Analytics: once you have some clean data, you can ask whatever question and it builds a dashboard on demand, helping users narrow the aperture on what\u2019s valuable to look at.
The problem originates from my previous experience at a BIG 4 :
Every deal starts with a data room: General ledgers. Accounting records. Contracts. Financial statements & Endless spreadsheets.
And before anyone can understand operate on top, a loot of offshore teams in poor countries have to reconstruct it by hand and inject it into a dashboard.
Does this kill Hebbia? No. But I think the winning platform for high-stakes financial work will need to do more than RAG and knowledge graphs.","title":"Show HN: SecondState \u2013 what if Hebbia started with raw ERP data?","updated_at":"2026-08-12T07:35:13Z","url":"https://trysecondstate.com/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"ayushpai"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["hebbia"],"value":"MLPs Are Hebbian Memories: A Simple Recipe for Fact-Storing Transformers"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["hebbia"],"value":"https://hazyresearch.stanford.edu/blog/2026-07-22-mlps-are-hebbians"}},"_tags":["story","author_ayushpai","story_49015247"],"author":"ayushpai","created_at":"2026-07-23T00:15:04Z","created_at_i":1784765704,"num_comments":0,"objectID":"49015247","points":2,"story_id":49015247,"title":"MLPs Are Hebbian Memories: A Simple Recipe for Fact-Storing Transformers","updated_at":"2026-07-23T00:19:49Z","url":"https://hazyresearch.stanford.edu/blog/2026-07-22-mlps-are-hebbians"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"ramon156"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["hebbia"],"value":"Hebbian Learning from First Principles (2024)"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://arxiv.org/abs/2401.07110"}},"_tags":["story","author_ramon156","story_48457922"],"author":"ramon156","created_at":"2026-06-09T07:43:23Z","created_at_i":1780991003,"num_comments":0,"objectID":"48457922","points":2,"story_id":48457922,"title":"Hebbian Learning from First Principles (2024)","updated_at":"2026-06-09T07:50:34Z","url":"https://arxiv.org/abs/2401.07110"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"skogstokig"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["hebbia"],"value":"Hebbia Imprint"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["hebbia"],"value":"https://design.hebbia.com"}},"_tags":["story","author_skogstokig","story_48383792"],"author":"skogstokig","created_at":"2026-06-03T13:30:59Z","created_at_i":1780493459,"num_comments":0,"objectID":"48383792","points":2,"story_id":48383792,"title":"Hebbia Imprint","updated_at":"2026-06-03T17:10:58Z","url":"https://design.hebbia.com"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"PaulHoule"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["hebbia"],"value":"Activity Propagation with Hebbian Learning"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://www.nature.com/articles/s42005-026-02638-z"}},"_tags":["story","author_PaulHoule","story_48122443"],"author":"PaulHoule","created_at":"2026-05-13T14:31:04Z","created_at_i":1778682664,"num_comments":0,"objectID":"48122443","points":2,"story_id":48122443,"title":"Activity Propagation with Hebbian Learning","updated_at":"2026-05-13T14:52:06Z","url":"https://www.nature.com/articles/s42005-026-02638-z"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"nixloop"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["hebbia"],"value":"Hebbia acquires FlashDocs \u2013 the #1 API to automate slide decks"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["hebbia"],"value":"https://www.upstartsmedia.com/p/exclusive-hebbia-acquires-flashdocs"}},"_tags":["story","author_nixloop","story_44433873"],"author":"nixloop","created_at":"2025-07-01T13:44:47Z","created_at_i":1751377487,"num_comments":0,"objectID":"44433873","points":2,"story_id":44433873,"title":"Hebbia acquires FlashDocs \u2013 the #1 API to automate slide decks","updated_at":"2025-07-01T13:46:35Z","url":"https://www.upstartsmedia.com/p/exclusive-hebbia-acquires-flashdocs"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"thebasedcapital"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["hebbia"],"value":"Show HN: BrainBox \u2013 Hebbian memory for AI coding agents"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/thebasedcapital/brainbox"}},"_tags":["story","author_thebasedcapital","story_47067820","show_hn"],"author":"thebasedcapital","children":[47067834],"created_at":"2026-02-18T23:26:07Z","created_at_i":1771457167,"num_comments":1,"objectID":"47067820","points":1,"story_id":47067820,"title":"Show HN: BrainBox \u2013 Hebbian memory for AI coding agents","updated_at":"2026-03-05T23:35:25Z","url":"https://github.com/thebasedcapital/brainbox"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"Opus_Warrior"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["hebbia"],"value":"Memory Infrastructure for AI Systems \u2013 Cascade, PyTorch Memory, Hebbian Mind"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://cipscorps.io/#"}},"_tags":["story","author_Opus_Warrior","story_46941508"],"author":"Opus_Warrior","children":[46941509],"created_at":"2026-02-09T04:13:34Z","created_at_i":1770610414,"num_comments":1,"objectID":"46941508","points":1,"story_id":46941508,"title":"Memory Infrastructure for AI Systems \u2013 Cascade, PyTorch Memory, Hebbian Mind","updated_at":"2026-03-05T23:32:54Z","url":"https://cipscorps.io/#"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"version_five"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["hebbia"],"value":"Hebbia raises $30M to launch an AI-powered document search tool"},"url":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["hebbia"],"value":"https://techcrunch.com/2022/09/07/hebbia-raises-30m-to-launch-an-ai-powered-document-search-tool/"}},"_tags":["story","author_version_five","story_32758946"],"author":"version_five","children":[32880789],"created_at":"2022-09-07T23:50:07Z","created_at_i":1662594607,"num_comments":1,"objectID":"32758946","points":1,"story_id":32758946,"title":"Hebbia raises $30M to launch an AI-powered document search tool","updated_at":"2024-09-20T12:02:37Z","url":"https://techcrunch.com/2022/09/07/hebbia-raises-30m-to-launch-an-ai-powered-document-search-tool/"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"vivan_121"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["hebbia"],"value":"Show HN: An open-source Rogo/Hebbia alternative I built for a PE client on AWS"},"url":{"matchLevel":"none","matchedWords":[],"value":"https://github.com/vivan1211/opengp"}},"_tags":["story","author_vivan_121","story_49255283","show_hn"],"author":"vivan_121","children":[49255284],"created_at":"2026-08-11T09:14:45Z","created_at_i":1786439685,"num_comments":0,"objectID":"49255283","points":1,"story_id":49255283,"title":"Show HN: An open-source Rogo/Hebbia alternative I built for a PE client on AWS","updated_at":"2026-08-11T09:16:56Z","url":"https://github.com/vivan1211/opengp"},{"_highlightResult":{"author":{"matchLevel":"none","matchedWords":[],"value":"rusanovych"},"story_text":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["hebbia"],"value":"I've been working on a hypothesis and want to get feedback from people who know more than I do.\nThe hypothesis\nIntelligence might be a phase transition at scale, not an algorithmic problem.\nFly: 100k neurons \u2014 no generalization\nMouse: 70M \u2014 basic associative learning\nHuman: 86B \u2014 abstract reasoning\nThis doesn't look like a smooth curve. It looks like thresholds. If that's true, then no amount of architectural cleverness crosses it \u2014 only scale + grounding does.\nThe grounding problem\nLLMs learn statistical distributions over text. "Apple" = token pattern. In biological systems "apple" = weight, texture, smell, hunger. Concepts without physical roots have a ceiling we're already approaching.\nThe architecture\nSphere topology: recurrent graph, no fixed signal direction, no enforced hierarchy\nHebbian learning only \u2014 no backprop\nDopamine-modulated consolidation with sleep/wake cycle\nSingle network: language + vision + motor through shared weights\nLateral inhibition + capacitor adaptation for stability \u2014 pure analog, already in Loihi\nGrounded in physical simulation, not text\nPrediction doesn't need to be engineered. Hebbian learning + physical grounding + continuous input should produce anticipation as an emergent property \u2014 same way it works biologically.\nWhy now\nAll components exist in current neuromorphic hardware. Full human scale = ~10,750 Loihi 3 chips, ~$150-200M. Below this threshold it probably won't work \u2014 that's the hypothesis, not a bug.\nThis needs real funding. But the architecture is ready to be attempted.\nWhat I'm looking for\nHas anyone attempted sphere topology on neuromorphic hardware? Is there prior work on Hebbian-only learning at this scale? Where does this obviously break?"},"title":{"fullyHighlighted":false,"matchLevel":"full","matchedWords":["hebbia"],"value":"Neuromorphic sphere topology Hebbian learning as a path to grounded intelligence"}},"_tags":["story","author_rusanovych","story_47331897","ask_hn"],"author":"rusanovych","created_at":"2026-03-11T05:04:17Z","created_at_i":1773205457,"num_comments":0,"objectID":"47331897","points":1,"story_id":47331897,"story_text":"I've been working on a hypothesis and want to get feedback from people who know more than I do.\nThe hypothesis\nIntelligence might be a phase transition at scale, not an algorithmic problem.\nFly: 100k neurons \u2014 no generalization\nMouse: 70M \u2014 basic associative learning\nHuman: 86B \u2014 abstract reasoning\nThis doesn't look like a smooth curve. It looks like thresholds. If that's true, then no amount of architectural cleverness crosses it \u2014 only scale + grounding does.\nThe grounding problem\nLLMs learn statistical distributions over text. "Apple" = token pattern. In biological systems "apple" = weight, texture, smell, hunger. Concepts without physical roots have a ceiling we're already approaching.\nThe architecture\nSphere topology: recurrent graph, no fixed signal direction, no enforced hierarchy\nHebbian learning only \u2014 no backprop\nDopamine-modulated consolidation with sleep/wake cycle\nSingle network: language + vision + motor through shared weights\nLateral inhibition + capacitor adaptation for stability \u2014 pure analog, already in Loihi\nGrounded in physical simulation, not text\nPrediction doesn't need to be engineered. Hebbian learning + physical grounding + continuous input should produce anticipation as an emergent property \u2014 same way it works biologically.\nWhy now\nAll components exist in current neuromorphic hardware. Full human scale = ~10,750 Loihi 3 chips, ~$150-200M. Below this threshold it probably won't work \u2014 that's the hypothesis, not a bug.\nThis needs real funding. But the architecture is ready to be attempted.\nWhat I'm looking for\nHas anyone attempted sphere topology on neuromorphic hardware? Is there prior work on Hebbian-only learning at this scale? Where does this obviously break?","title":"Neuromorphic sphere topology Hebbian learning as a path to grounded intelligence","updated_at":"2026-03-11T05:05:45Z"}],"hitsPerPage":20,"nbHits":282,"nbPages":15,"page":0,"params":"query=hebbia&advancedSyntax=true&analyticsTags=backend","processingTimeMS":148,"processingTimingsMS":{"_request":{"roundTrip":23},"afterFetch":{"format":{"highlighting":2,"total":2},"merge":{"mergeLoop":{"prepareNextHit":34,"total":34},"total":35},"total":35},"fetch":{"query":3,"total":4},"getIdx":{"load":{"gens":98,"synonyms":9,"total":108},"total":108},"total":149},"query":"hebbia","serverTimeMS":244}