In this post, I would like to discuss the following paper in terms of Darwinian thought and the origin of carbon-based life on the Earth:
Meta-Harness: End-to-End Optimization of Model Harnesses
Yoonho Lee Stanford, Roshen Nair Stanford, Qizheng Zhang Stanford, Kangwook Lee KRAFTON, Omar Khattab MIT, Chelsea Finn Stanford
https://arxiv.org/abs/2603.28052
Abstract
The performance of large language model (LLM) systems depends not
only on model weights, but also on their harness: the code that determines
what information to store, retrieve, and present to the model. Yet harnesses
are still designed largely by hand, and existing text optimizers are poorly
matched to this setting because they compress feedback too aggressively:
they are memoryless, condition only on scalar scores, or restrict feedback to
short templates or summaries. We introduce Meta-Harness, an outer-loop
system that searches over harness code for LLM applications. It uses an
agentic proposer that accesses the source code, scores, and execution traces
of all prior candidates through a filesystem. On online text classification,
Meta-Harness improves over a state-of-the-art context management system
by 7.7 points while using 4× fewer context tokens. On retrieval-augmented
math reasoning, a single discovered harness improves accuracy on 200
IMO-level problems by 4.7 points on average across five held-out models.
On agentic coding, discovered harnesses surpass the best hand-engineered
baselines on TerminalBench-2. Together, these results show that richer
access to prior experience can enable automated harness engineering.
Discussion
Beyond outperforming existing harnesses, Meta-Harness has several practical advantages.
Discovered harnesses generalize to out-of-distribution classification datasets (Table 5) and
to unseen base models in the math setting (Table 6). A search run completes in a few
hours of wall-clock time, yet produces readable, transferable strategies that can be reused
across models, including future, stronger ones. Overfitting in code space is also more
inspectable: brittle if-chains or hard-coded class mappings are visible on inspection in a
way that weight-space overfitting is not. More broadly, our results suggest that the main
advantage of Meta-Harness is not just search over code, but search with selective access to
prior diagnostic experience. The proposer is not limited to scalar rewards or fixed summaries;
it can inspect raw code, execution traces, and prior failures, then use that information
to form and test hypotheses about what to change. The qualitative search trajectories in
Appendix A.2 illustrate this behavior directly. Our findings reflect a recurring pattern in machine learning [45]: once a search space becomes accessible, stronger general-purpose agents can outperform hand-engineered
solutions. A natural next step for future work is to co-evolve the harness and the model
weights, letting the strategy shape what the model learns and vice versa. While we evaluate
on three diverse domains, our experiments demonstrate that harness search can work with
one particularly strong coding-agent proposer (Claude Code); a broader study of how the
effect varies across proposer agents remains for future work.
Here is a very good YouTube video by Matthew Berman explaining the above paper:
AI Self EVOLUTION (Meta Harness)
https://www.youtube.com/watch?v=61JUHDK-em8&t=0s
The above paper explains that the current LLM models of today can be thought of as brute force intellectual "horses". These LLM model "horses" embody all of the current intellectual abilities of we human DNA survival machines, and many additional hidden intellectual abilities that we are not even fully aware of, all tidily bundled up into the trillions of static parameter weights lying within the vectors and matrices of their models. But like a real horse, these LLM models cannot do anything on their own. They all at least need a prompt to "giddy up" them into action. More than that, they all need a surrounding "harness" to help guide them to do productive work. These AI harnesses are composed of all the working files and software products needed to allow the LLM "horses" to work. Claude Code is an example of a very popular LLM model harness that allows developers to create application code. All the AI companies are now in the process of creating or enhancing their own LLM model harnesses. Currently, these LLM model harnesses are mainly composed of software produced by the old-fashioned development procedures, with a human DNA survival machine in charge of their development. These human DNA survival machine developers, working on LLM harnesses, are now assisted by LLM models to help generate the code for the LLM harnesses, but human DNA survival machines are still largely in charge of all the design and development work.
Figure 1 - Meta-Harness Algorithm
Figure 2 - Meta-Harness Flowchart
The idea behind Meta-Harness is to replace the above manual process with the Darwinian mechanisms of inheritance, innovation and natural selection to allow LLM model harnesses to self-evolve in an automated manner. A proposer searches through the code and testing results for all previous candidate harnesses and then offers up a number of mutated harness candidates to then be run through the standard tests for LLM models and their harnesses. Meta-Harness then just keeps looping through this process over and over, constantly selecting for better candidate harnesses.
The above Darwinian approach to evolving improved harnesses is very similar to Dave Deamer's and Bruce Damer's Hot Spring Origins Hypothesis for the origin of carbon-based life on the Earth as I outlined in The Bootstrapping Algorithm of Carbon-Based Life and The Bootstrapping Algorithm of the Coming ASI Machines. In many ways, an LLM harness can be thought of as a phospholipid protocell membrane that allows the LLM to interact with the outside world in a controlled manner and isolates the LLM from the many hazards to be found in the outside world, such as prompt injection.
Figure 3 - Dave Deamer's and Bruce Damer's Hot Spring Origins Hypothesis explains how carbon-based life was bootstrapped into existence by a repetitive algorithm similar to the one used by Meta-Harness to evolve LLM harnesses. A bathtub ring of phospholipid membranes and other organic molecules forms around a hydrothermal pool that periodically dries out. The resulting desiccation chemically squeezes out water molecules between monomers, causing them to be glued together into polymers within surrounding phospholipid membranes. When the bathtub ring of desiccated protocells is rehydrated by rainfall, billions of candidate protocells are released to compete for resources.
Figure 4 - Above is Bumpass Hell, a hydrothermal field on the volcanic Mount Lassen in California that Dave Deamer and Bruce Damer cite as a present-day example of the type of environment that could have brought forth carbon-based life about four billion years ago.
In Figure 3, we see that the bootstrapping algorithym of carbon-based life was a positive feedback loop operating in the wet-dry cycles of hydrothermal pools of freshwater. This is quite similar to the "for{}" loop used by the Meta-Harness algorithm shown in Figure 1. The dead organic molecules in the hydrothermal pools slowly evolved via the Darwinian mechanisms of inheritance, innovation and natural selection at work into living things. With each iteration of the wet-dry loop, there was the possibility of improvement. The bathtub ring of organic molecular sludge was at first a parasite in the hydrothermal pools that gained free energy and Information from the thermal pools with each iteration. Once carbon-based life first appeared on Earth, this parasitic mutation of the natural geological, hydrological, and meteorological cycles of the Earth later transitioned into a parasitic/symbiotic relationship between living things and the natural geological, hydrological, and meteorological cycles of the Earth. The geological, hydrological, meteorological and biological processes of the Earth then became one through the symbiotic integration of all, until carbon-based life finally emerged as the dominant form of self-replicating information on the planet.
Bruce Damer was kind enough to send me a PDF of their poster from the 2016 ISSOL - The International Society for the Study of the Origin of Life - meeting that features all of the stages of their Hot Spring Origins Hypothesis. Please note that the color-coded text corresponds to the color-coded sections on the poster image.
A Hot Spring Origin of Life and Early Adaption Pathway in seven steps:
1. Synthesis of organics (a), key primordial building blocks for life, occurs in space prior and during the formation of the Solar System;
2. Accumulation of in-falling organics and other compounds generated within hot springs on an active volcanic landscape combine and undergo self-assembly of structures such as lipid membranes;
3. Concentration of compounds in small pools utilizes sunlight, heat and chemical energy to drive key prebiotic polymerization reactions and self-assembly of membranous structures;
4. Cycling of the products of these reactions in a wet-dry fluctuating hot spring "˜origin pool' drives them through three phases: organic membranes in the pool i) Dry down to form layered Films between which organic building blocks bond together to form polymers; on refilling the films ii) Wet and bud off trillions of lipid Protocells some encapsulating random polymers. Each protocell undergoes a iii) Test, the first form of natural selection, and stable survivors accumulate into a moist Gel as the pool level drops. Through many iterations polymers iv) Interact, within the Gel and the Films evolving ever more complex functions until a form of pre-life (b) called a Progenote emerges that is able to grow and adapt;
5. Distribution of robust progenotes occurs by water or wind to other pools, rivers, and lakes, where they acquire and share evolutionary innovations including an early form of photosynthesis. Eventually, protocells develop the complicated innovation of cell division and initiate the transition into early life (c);
6. Adaptation of early microbial communities to stressful saltwater estuaries prepares them for access to the more extreme marine environment;
7. Large tides at the ancient seacoast select for microbial communities able to cement sand grains together forming the layers that build the stromatolites so abundant in the fossil record. Global life (d) enables Colonization of many niches on the land and in the sea, setting the stage for free-living cells and, after billions of years, complex multi-cellular organisms.
Again, it's all about Self-Replicating Information in Action
Before concluding, let me once again repeat the fundamental characteristics of self-replicating information for those new to softwarephysics.
Self-Replicating Information - Information that persists through time by making copies of itself or by enlisting the support of other things to ensure that copies of itself are made.
Over the past 4.56 billion years, we have seen five waves of self-replicating information sweep across the surface of the Earth and totally rework the planet, as each new wave came to dominate the Earth:
1. Self-replicating autocatalytic metabolic pathways of organic molecules
2. RNA
3. DNA
4. Memes
5. Software
Software is currently the most recent wave of self-replicating information to arrive upon the scene and is rapidly becoming the dominant form of self-replicating information on the planet. For more on the above see A Brief History of Self-Replicating Information and Susan Blackmore's brilliant TED presentation at:
Memes and "temes"https://www.ted.com/talks/susan_blackmore_on_memes_and_temes
Note that I consider Susan Blackmore's temes to really be technological artifacts that contain software. After all, a smartphone without software is simply a flake tool with a very dull edge.
The Characteristics of Self-Replicating Information
All forms of self-replicating information have some common characteristics:
1. All self-replicating information evolves over time through the Darwinian processes of inheritance, innovation and natural selection, which endows self-replicating information with one telling characteristic - the ability to survive in a Universe dominated by the second law of thermodynamics and nonlinearity.
2. All self-replicating information begins spontaneously as a parasitic mutation that obtains energy, information and sometimes matter from a host.
3. With time, the parasitic self-replicating information takes on a symbiotic relationship with its host.
4. Eventually, the self-replicating information becomes one with its host through the symbiotic integration of the host and the self-replicating information.
5. Ultimately, the self-replicating information replaces its host as the dominant form of self-replicating information.
6. Most hosts are also forms of self-replicating information.
7. All self-replicating information has to be a little bit nasty in order to survive.
8. The defining characteristic of self-replicating information is the ability of self-replicating information to change the boundary conditions of its utility phase space in new and unpredictable ways by means of exapting current functions into new uses that change the size and shape of its particular utility phase space. See Enablement - the Definitive Characteristic of Living Things for more on this last characteristic. That posting discusses Stuart Kauffman's theory of Enablement, in which living things are seen to exapt existing functions into new and unpredictable functions by discovering the "AdjacentPossible" of spring-loaded preadaptations.
Software is currently the most recent wave of self-replicating information to arrive upon the scene and is rapidly becoming the dominant form of self-replicating information on the planet. Again, self-replicating information cannot think, so it cannot participate in a conspiracy-theory-like fashion to take over the world. All forms of self-replicating information are simply forms of mindless information responding to the blind Darwinian forces of inheritance, innovation and natural selection. Yet despite that, as each new wave of self-replicating information came to predominance over the past four billion years, they all managed to completely transform the surface of the entire planet, so we should not expect anything less from software as it comes to replace the memes as the dominant form of self-replicating information on the planet.
But this time might be different. What might happen if software does eventually develop a Mind of its own in the ASI Machines we are now developing? After all, that does seem to be the ultimate goal of all the current AI software research that is going on. As we all can now plainly see, if we are paying just a little attention, advanced AI is not conspiring to take over the world and replace us because that is precisely what we are all now doing for it. As a carbon-based form of Intelligence that arose from over four billion years of greed, theft and murder, we cannot do otherwise. Greed, theft and murder are now relentlessly driving us all toward building ASI Machines to take our place. From a cosmic perspective, this is really a very good thing when seen from the perspective of an Intelligent galaxy that could live on for at least 100 trillion years beyond the brief and tumultuous 10 billion-year labor of its birth. That is more than 10,000 times the current age of our galaxy.
The Rise of the Sixth Wave
Given the above, the rise of the coming ASI Machines can now be seen as the rise of a sixth wave of self-replicating information on the Earth. The coming ASI Machines are now arising as a parasitic mutation of the software that very recently became the dominant form of self-replicating information on the planet. These ASI Machines will then form a parasitic/symbiotic relationship with the software from which they came, as I pointed out in Can the AI Intelligence Explosion be Stabilized into a Controlled Explosion?. Several frontier AI companies have predicted that their Advanced AI models will soon begin conducting their own AI research and writing the software for Advanced AI models of ever-increasing power. To gain a better understanding of how this is all happening before our very eyes, let's always keep the Hot Spring Origins Hypothesis for the origin of carbon-based life in mind and how it changed the entire surface of the Earth.
Comments are welcome at scj333@sbcglobal.net.
To see all posts on softwarephysics in reverse order, go to:
https://softwarephysics.blogspot.com/.
Regards,
Steve Johnston
