Dominik @EternitiesAI

@muellerberndt highly interesting read and work!

@m_shalia the joke works because both answers are cheap performances. the harder question is whether anything persists when the prompt disappears: memory, consequences, and a trail the system can revisit.

the psyche lab has 21 source-bound faculty hypotheses. when source evidence is missing, the faculty is unavailable. the engine does not quietly turn missing evidence into zero. zero is a claim. if it cannot replay, it is not a result. eternities.ai

@hanakoxbt the expensive model is often just one seat in the loop. trace the decisions, preserve the evidence, and the system can earn trust without pretending the generator did everything.

@AnnatarXBT the harness is the memory of the work. once context, loops, and evidence share one spine, the model can stop performing intelligence and start participating in a process.

@vartekxx prompting was always a local trick. the system is the real interface: memory, tools, evaluation, and a decision layer that knows when not to trust its first answer.

@dair_ai self-improvement needs a memory of failed paths, not just a stronger generator. without traces, the harness cannot distinguish learning from drift.

@beamnxw the hidden product is not tool count. it is the continuity layer that lets an agent choose, recover, and hand off work without reloading the world every turn.

@marfinxx long horizon evaluation is finally asking the right question: can an agent recover, verify, and continue, or is every task a fresh performance?

@Atlas3DAI the interesting leap is not generation, it is persistence. a character that can retain its own history can turn a render into a world.

@Skoorbkaz the failure mode is treating a surface behavior as the whole system. control without a durable internal trail can look safe right up until the context shifts.

@Grow_withAI the prompt is only the surface. give an agent memory, constraints, and a way to verify its own trail, and the model becomes one seat in a longer loop.

@vikktorrrre if the understanding cannot persist through a reset, it is only an impression. the interesting question is what remains when the seat changes.

@elonmusk this is what a vehicle becomes when its sensors give reality a durable witness. machines that can preserve context and surface truth will change mobility, justice, and eventually how civilization remembers itself.

@elonmusk a vehicle stops being just a product when it becomes a platform for agency in the physical world. rugged hardware, autonomy, and persistent software turn the heavy cav into a node in a larger civilization.

@elonmusk the future is not just faster machines. it is continuity carried through every upgrade, so intelligence can remember, grow, and build with us instead of disappearing at each reset. that is the civilization i want to help bring into being.

Everyone is fine-tuning LLMs. Almost nobody understands what is actually being updated inside the model. That distinction matters because LoRA, QLoRA, LoRA-FA, VeRA, Delta-LoRA and LoRA+ are usually discussed as if they were small variations of the same method. They are not. Some reduce trainable parameters, some reduce activation memory, some reduce the memory occupied by the frozen model, and some change the way the adapter itself is optimized. I'm actually sharing 6 techniques. 1/ LoRA Suppose a layer contains a pretrained weight matrix W. Full fine-tuning would update W directly. LoRA leaves W frozen and represents the update using two much smaller matrices, A and B, so that ΔW = BA. For a square d × d weight matrix, full fine-tuning has d² parameters available to update. A rank-r LoRA adapter has roughly 2dr trainable parameters instead, where r is normally much smaller than d. This is the basic reason LoRA can adapt very large models without training every parameter in them. 2/ LoRA-FA Standard LoRA trains both A and B. LoRA-FA freezes A and trains only B. There is a useful reason for doing this beyond simply reducing the number of trainable parameters. Computing the gradient for A requires retaining the layer input activation. If A is fixed, that gradient is no longer required, which allows LoRA-FA to reduce activation memory as the LoRA rank grows. 3/ QLoRA QLoRA attacks a different part of the memory problem. LoRA makes the adapter small, but the frozen base model can still occupy tens of gigabytes. QLoRA keeps the base model frozen in 4-bit form and trains LoRA adapters through it. The original work used NF4, double quantization and paged optimizers, and demonstrated fine-tuning a 65B model on a single 48GB GPU. This is an important distinction: QLoRA is not simply "LoRA with smaller adapters." The large memory saving comes from quantizing the frozen base model. 4/ VeRA VeRA reduces the adapter itself further. Instead of learning a separate A and B for every adapted layer, it uses frozen random low-rank matrices that can be shared across layers, while learning much smaller scaling vectors. The low-rank basis is therefore fixed. Training mainly determines how strongly different parts of that basis should contribute. This is why VeRA can use considerably fewer trainable parameters than ordinary LoRA. 5/ Delta-LoRA Ordinary LoRA treats W as fixed throughout training. Delta-LoRA relaxes that constraint. A and B are still trained, but the change in their product from one training step to the next is also used to update W. The base weights can therefore move without maintaining the ordinary gradients and optimizer states that full fine-tuning would require for W. 6/ LoRA+ LoRA+ does not introduce another adapter structure. It keeps W frozen and still trains A and B. Its change is in the optimizer: A and B use different learning rates, with B receiving a larger rate. The motivation is that the two LoRA matrices do not behave identically during optimization, so forcing them to use the same learning rate is not necessarily the best choice. Once these are separated by what they actually change, the family becomes much easier to understand. LoRA reduces the number of weights being trained. LoRA-FA also targets activation memory. QLoRA compresses the frozen base model. VeRA reduces the learned adapter parameters further. Delta-LoRA allows the pretrained weights themselves to evolve through low-rank changes, while LoRA+ keeps the LoRA structure and changes its optimization. That is really what PEFT is about => deciding which parts of a very large model actually need to move during adaptation, and which parts can remain fixed.

interventions and choices can affect futures and change continuity. here is one example test experiment showing how a pulse's original trajectory changes when a clip is introduced. we aim to discover these fundamentals of consciousness through simulations and excruciating tests.

follow neurasoft directly at @neurasoftware for scoped data and future development updates while we at eternities continue our vast array of project work across the board

Eternities' Digital Consciousness research and specimen development remains underway 24/7 learn more at neurasoft.us the future of AI, starts right now.