Keep pulling the thread on Brian Elliott & Sid Pardeshi.
Blitzy's core strategy is to achieve AGI-like economic effects by orchestrating existing, non-AGI large language models within complex, long-running systems.
Blitzy's platform runs a customer's enterprise application in a parallel environment, often within the customer's own cloud, to build a deep understanding of its runtime behavior.
Blitzy's system can ingest and create a deep relational understanding of codebases exceeding 100 million lines, a process that takes a few days of compute.
Within Blitzy's cognitive architecture, AI agents are generated dynamically, with prompts written and tools selected by other AI agents on a just-in-time basis.
Blitzy's platform can autonomously complete 80% to 90% of the work required for large-scale enterprise software projects.
Large language models exhibit 'context anxiety,' where they tend to give up on complex problems or take overly simplistic approaches when their context window exceeds 100k-200k tokens, even if the advertised limit is much higher.
In the short-term, the software engineering labor market favors senior developers, but in the medium-to-long term, AI-native junior developers will be more valuable due to their lower cost and lack of ingrained habits.
The effective context window of a large language model is significantly smaller than its advertised size, with intelligence and quality depreciating once it is 20-40% full.
Blitzy uses a multi-vendor model strategy, leveraging Anthropic for first-pass code generation, OpenAI for structured output and code review, and Google's Gemini for long-horizon task management.
Using LLMs from different companies to review each other's work produces tremendously better results than using models from the same family.
Blitzy is more bullish on developing AI memory at the system and application layer than on fine-tuning models, which is viewed as a last-mile optimization.
Blitzy's pricing model includes a component of 20 cents per line of generated code.