Why does the CEO of TypeSafe argue that code models have to give up humanisation?

This article is part of the PRO Members Communication, and follows "PRO Members of Machine Heart" at the end of the article to read more on the topic.

With the deep penetration of generative AI and codebase models, code agents and programming systems are gradually replacing humans as the primary consuming subjects of APIs. Traditional AI models consume huge amounts of computing power and context-based budgets in natural language color, customizable responses and complex layouts. This personalization expression, while causing a waste of computing power, also enables developers to experience delays and uncertainties when faced with high demand for advanced and powerful industrial-level code. As a former OpenAI core researcher, InstructGPT and ChatGPT co-author, the main developer of the machine-generated model Jev, TypeSafe CEO and founder Diogo Almeida proposed that the standard for evaluating codebase models needs to be rigorously aligned with the US dollar, alongside smart patches, to completely eliminate the costs of human software engineering, and thus drive the development of automated software and automated code systems by directly eliminating biases and eliminating the cost of each RCD.

Directory

01. Why「RLHF」Does the route mislead code developers?
What is "smart per dollar" and "pareto front"?...

Why does the RLHF route mislead code developers?

1 Diogo Almeida, co-founder and CEO of TypeSafe AI, recently interviewed around the release of the machine native model Jev, which dissects the optimization misconceptions of traditional text models in the code scenario, and suggests that the code model must move away from a personalized alignment and build a high-reliability underlying computing infrastructure around every dollar of intelligence and programming on a global scale.

1 Diogo Almeida was a core researcher for the OpenAI language team, involved in the research and development of InstructGPT and ChatGPT, and is currently co-founder and CEO of TypeSafe AI.

2 Jev is the native model of the System 1 machine introduced by TypeSafe AI, which abandoned text generation in line with human preferences and did not produce free text, designed for high concurrency and strong code decision-making scenarios. Following the launch of Jev, it quickly attracted large numbers of developers and platforms to the production environment with relatively considerable latency and cost performance.

2 Almeida believes that the structural transformation of API consumer subjects is reversing model evaluation standards, and that past text optimization to accommodate human reading preferences will result in serious computational waste when key callers transition from humans to program code and Agent.

1 In the past few years, the AI industry has been caught up in the mentality of humanizing models, with mainstream manufacturers spending huge amounts of computing power with context-based budgets on natural language coloring, custom responses and complex layouts; but in the code scene, the output received is by downstream compilers and analyzers, which require extremely low response latency and extremely high structural certainty.

2 In Parretto's avant-garde engineering estimates, there is a clear competitive relationship between natural language expressions and logical validation of arithmetic, completely abandoning the excess natural language expressions that allow each unit of arithmetic and the generation of tokens to be put into a type iteration to achieve a quantitative increase in system throughput.

3 Around the redesigned engineering architecture, Almeida proposes that per-dollar intelligence is the core indicator for measuring the efficiency of the industrial-grade code infrastructure. This indicator requires the optimization goal of the model to move away from the fallacy of simply pursuing benchmark test scores and instead concentrate all of the engineering calculations on improving unit capital investment ratios and effective reasoning ability.

1 Per dollar smart refers to the total number of definiteness logics and types of security codes the model can stabilize for delivery at a specific API Token cost and computational budget constraints.

In model training and alignment pathways, Almeida points out that the RLHF paradigm of reinforcement learning based on human feedback has problems with model collapse, mode collapse, and calibration failures, which may cause models to over-supply to human preferences, decrease in diversity, and confident but erroneous output.

1 The RLHF is designed to align human communication tone with secure rejection responses, but the technology gives models an over-explanatory tendency to inject uncertainty into the generated code, resulting in confusion of uncertainty in explanatory text and format noise, constituting an abnormal source in code analysis.

2 This confusion and delusion of natural language created to accommodate human preferences undermines the syntactical certainty required for downstream machine calling scenarios, leading to unexpected collapses of models when dealing with automated compilation and structured interface resolution.

How is code reinforcement learning moving towards the RLCD paradigm?

1 As big models move deeper into the field of complex reasoning (such as mathematics and code) and RLVR (enhanced learning based on verifiable results) gradually rise, Almeida believes that while RLVR can solve specific problems such as Navier-Stokes, it will exacerbate the phenomenon of "jagged intelligence"...