[News Brief] The $100k Checkpoint, The Legacy OCR Fix, and The Antigravity Reality Check
Why the "agentic harness" matters more than pre-training, how Landing AI solved the dirty PDF crisis, and why Windsurf remains the superior IDE.
This week’s session was a blend of high-level strategic debate and deep-dive technical investigation. We dissected Amazon’s expensive new training tool, launched a new internal support structure, and scrutinized the battle between Google’s Antigravity and Windsurf.
However, the highlight was a practical task force investigation into solving a nagging enterprise problem: Legacy OCR.
Here are the spiky points of view and technical findings from the session.
The Economics of Pre-Training: Amazon Nova Forge
We discussed Amazon’s release of Nova Forge, a platform designed to streamline the training of custom Frontier models. The promise is enticing: by investing in pre-training rather than just fine-tuning, companies can avoid “catastrophic forgetting” and hallucinations.
However, the team expressed skepticism regarding the value proposition for standard SaaS applications.
Spiky Point of View: Renting vs. Owning
Leonardo Gonzalez argued that for the $100,000 enterprise license (which covers access but not compute), a company should own the model weights and be able to serve them anywhere.
Nova Forge essentially charges a premium for access to a walled garden.
In contrast, open initiatives like OLMo 3 provide the full pipeline—checkpoints, data, and recipes—allowing for custom pre-training at a fraction of the cost.
The Strategic Shift: The Harness Matters More
The consensus was that for most use cases, the LLM itself is becoming a commodity. The true differentiator is no longer the model, but the “agentic harness”—the logic and orchestration layer wrapped around the model.
“The value addition is largely on the agentic harness now... assuming that the LLMs have been more or less commoditized.” — Praveen Koka
The Task Force: Solving Legacy OCR
A major friction point for the Account Management team has been mining data from old, scanned contracts (often from the 1990s). We investigated four specific tools to solve this: PyTesseract, Google Cloud Document AI, Gemini Chat, and Landing AI.
The Findings:
PyTesseract: Too rudimentary. It produces text with defects that require a heavy LLM cleanup layer.
Gemini Chat: A strong contender. It performs better than ChatGPT on extraction, specifically if you prompt it for page numbers, clause numbers, and confidence scores.
The Winner: Landing AI.
Why Landing AI Won
Identified as the “best of the lot,” Landing AI (founded by Andrew Ng) offered a critical feature for enterprise trust: Visual Grounding. When the model answers a question, it highlights the exact chunk of text in the PDF source, allowing for instant human verification.
The IDE Battle: Antigravity is just a Locked-Down Windsurf
We revisited the buzz around Google’s Antigravity IDE. The team clarified a critical piece of context: Antigravity is essentially a fork of Windsurf, based on IP that Google licensed.
The Constraint
Antigravity is currently in a “free public preview” state. However, it suffers from severe usage limits. Once a user hits the limit, they are locked out regardless of their enterprise status or willingness to pay.
The Recommendation
Windsurf (now owned by Cognition) has evolved past the Google fork. It remains the superior choice for production because:
It is model-agnostic, supporting over 50 models including the newly released Claude Opus 4.5.
It features advanced code architecture maps that Antigravity currently lacks.
It does not artificially cap usage for paid users.
“I’m encouraging people to try Windsurf because it’s the same base product, without any of the restrictions and with just an enormous amount of models available.” — Leonardo Gonzalez
New Initiative: CoE Assist
We officially introduced AI Center of Excellence (CoE) Assist. This is a formal pipeline designed to help Business Units with:
Advice: Tool selection and model validation.
Architecture Review: Acting as a sounding board for your designs.
Research Assistance: Deep dives into specific topics.
POCs & Integration: Building proofs of concept for specific business problems.