{
  "video": {
    "id": "T-42NrYjwxM",
    "title": "Our Smaller Model Beat Opus on Finance Tasks - Alexander Shevchenko, Ramp",
    "duration": 855,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": ""
  },
  "analysis": {
    "video_id": "T-42NrYjwxM",
    "title": "Our Smaller Model Beat Opus on Finance Tasks - Alexander Shevchenko, Ramp",
    "one_liner": "Ramp keeps frontier models as orchestrators but hands financial question-answering to a post-trained Qwen 3.5 specialist that beat Opus 4.6 on exact task accuracy while running much faster, showing why an application-layer company still does applied research instead of waiting for the frontier.",
    "summary": "Alex from Ramp answers the question \"why not just wait?\" for better frontier models, even though Ramp's LLM use cases grew 123% in a year while unit economics improved as newer models arrived. He compares today with 2023, when Ramp shipped agents on GPT-4 rather than waiting for GPT-4.5 or GPT-5, and argues that teams can still dig a layer down to build better experiences. He presents three Ramp Labs cases: a post-trained specialist model, KV cache \"cartridges\", and agentic map-reduce swarms. The common pattern is saving time (latency) and money (token cost), and that research has also influenced Ramp's token spend management and Ramp Router products.",
    "key_points": [
      "Over the last year, Ramp's LLM use cases increased 123%; top features are growing in token spend but becoming cheaper on a unit-economic basis as cheaper, faster, better frontier models are released.",
      "In 2023, building agents on GPT-4 required strict guardrails, no backtracking or thinking tokens, sweating over prompts where synonyms could change behavior, and tightly scoped context. Ramp still shipped agents such as Ramp Tour Guide.",
      "Ramp Sheets launched in November as an early spreadsheet agent, before the two frontier labs launched theirs. The biggest friction with Opus 4.6 was the slow, expensive thinking spent planning and researching many dropped-in sheets when asked to build a DCF or model a financial scenario.",
      "With Prime Intellect, Ramp post-trained a Qwen 3.5 model into a financial question-answering and retrieval specialist that scored above Opus 4.6 on exact task accuracy and was much faster. It was sent to prod, with the frontier model kept as orchestrator.",
      "Cartridges, a family of KV cache techniques Ramp is still exploring, compress up to 20 to 30 times more context into the KV cache offline. For month-end close, one can be built overnight from the ledger, aging report and documents so the agent already has the context when the financial professional opens it on the first of the month.",
      "Cartridges are a much lower lift than post-training, but the labs might ship something similar in their APIs. Alex says Ramp would be happy with that, because the tribal knowledge it is building would let it hit the ground running on day one.",
      "Using its cloud agent Ramp Inspect for infrastructure, Ramp ran 10,000 parallel agents to pen test all of Ramp, then used LLMs to deduplicate tickets and adversarially reproduce them. It found and patched seven high-severity vulnerabilities, including a Lua script injection sendable via email that required understanding multiple components end-to-end.",
      "Open-weight models up to five times cheaper than frontier reasoning models do not get a five times lower detection rate when run as a swarm, so spending more on cheaper tokens can achieve almost as good a detection rate."
    ],
    "takeaways": [
      "Don't wait for the next frontier model to serve users; ship now and dig a layer below what off-the-shelf models and MCPs give you.",
      "Find where the frontier model has the most friction, such as slow, expensive planning and retrieval, and hand that step to a faster post-trained specialist while keeping the frontier model as orchestrator.",
      "For predictable recurring tasks like month-end close, prepare context offline before the user starts, for example with cartridges built overnight, instead of having the agent retrieve it live each time.",
      "For work like pen testing, try swarms of cheaper open-weight models with an LLM map-reduce step that deduplicates findings and adversarially reproduces them to confirm they are real and important.",
      "Invest in techniques the labs may later ship in their APIs anyway, because the internal know-how lets you use those features well from day one."
    ],
    "topics": [
      "applied research",
      "post-training",
      "specialist models",
      "kv cache",
      "agent swarms",
      "map-reduce",
      "security",
      "llm cost optimization"
    ],
    "tools": [
      "Ramp",
      "Ramp Labs",
      "Ramp Sheets",
      "Ramp Tour Guide",
      "Ramp Inspect",
      "Ramp Router",
      "Prime Intellect",
      "Qwen 3.5",
      "Opus 4.6",
      "GPT-4",
      "GPT-4.5",
      "GPT-5",
      "MCP",
      "Nvidia Switchyard",
      "Datadog"
    ],
    "quotes": [
      {
        "text": "we decided, well, we need to ship these things. We can't wait for GPT-4.5 or GPT-5 to come out.",
        "at": "02:20",
        "url": "https://www.youtube.com/watch?v=T-42NrYjwxM&t=140s"
      },
      {
        "text": "And even if they were to do it in the API, we would actually be happy",
        "at": "07:31",
        "url": "https://www.youtube.com/watch?v=T-42NrYjwxM&t=451s"
      },
      {
        "text": "And no single component of that stack was unsafe.",
        "at": "09:01",
        "url": "https://www.youtube.com/watch?v=T-42NrYjwxM&t=541s"
      },
      {
        "text": "that you can actually use open weight models that are up to five times cheaper",
        "at": "09:46",
        "url": "https://www.youtube.com/watch?v=T-42NrYjwxM&t=586s"
      }
    ],
    "words": 2684
  },
  "summary_url": "/#T-42NrYjwxM",
  "transcript": {
    "html": "/transcripts/T-42NrYjwxM.html",
    "txt": "/transcripts/T-42NrYjwxM.txt",
    "vtt": "/transcripts/T-42NrYjwxM.vtt"
  }
}