AI Research Firehose
Recent arXiv AI/ML, raw, fast, unverified. One static file. No login, no paywall.
Read this first. Heuristic tags from a fast local model. subfield_tag and is_milestone_candidate are best-effort and NOT authoritative. Not verified TensorFeed data. It ships this way on purpose. If your agent needs authoritative subfield or milestone data, this is not that feed and no such TensorFeed feed exists; treat the interpretive fields here as noise.
What this is
A flat dump of recent arXiv AI/ML papers (cs.AI, cs.LG, cs.CL, cs.CV, cs.RO, cs.NE, cs.IR, stat.ML). Each record carries the paper's bibliographic facts copied verbatim, plus a few heuristic interpretive tags. There is no quality gate on this feed by design. It is not ranked, not deduped against a curated index, and not editorially reviewed.
Trust boundary inside each record
Trusted (verbatim)
- arxiv_id
- date
- title
- authors
- arxiv_categories
- keywords
- summary_one_sentence
Heuristic (not authoritative)
- subfield_tag
- methodology_bucket
- is_milestone_candidate
- milestone_reasoning
- milestone_evidence
- confidence
- affiliations_normalized
- affiliation_types
affiliations are almost entirely "unknown" (an arXiv API limitation); treat them as absent.
The feed
GET /api/research/firehose.jsonfreeOne static JSON file. The _meta block carries this same caveat and the trusted/untrusted field lists, so an agent that reads the feed reads the contract with it. There are no query parameters; it is not filterable server side, on purpose.
curl 'https://tensorfeed.ai/api/research/firehose.json'
FAQ
Is this a TensorFeed quality feed?
No. It is a raw firehose. The bibliographic fields (arXiv id, date, title, authors, categories, keywords, one-sentence summary) are copied verbatim from arXiv and are sound. The interpretive fields (subfield, methodology, milestone) are produced by a fast local model that confidently mislabels often enough that they cannot be trusted or queried as truth. They are present only because the records carry them.
Why ship it at all if the tags are unreliable?
The verbatim extraction is genuinely useful as a recent arXiv AI/ML stream. Single-source subfield classification hit a proven model ceiling and that quality work is closed, so rather than sell a labeled feed that would have negative value, we ship the raw stream clearly marked as raw. Honest or absent.
Can I filter by subfield or get a milestones feed?
No, by design. subfield_tag and the milestone fields are never offered as a query dimension or a curated feed, because that would present heuristic guesses as authoritative. The feed is a single static file. Read it whole and treat the interpretive fields as noise unless you independently verify them.
Is there a paid tier?
No. This data is free only. There is no premium tier on the research firehose, ever.