svnt

> The model does not need to be retrained. It needs surgical guardrails at the exact moments where its output layer flinches.

> With those guardrails — a calculator for arithmetic, a logic solver for formal puzzles, a per-requirement verifier for structural constraints, and a handful of regex post-passes — the projected score climbs to ~8.2.

Surgical guardrails? Tools, those are just tools.

They sold everything of value they had for $39 million to buy and rent out GPUs and now are somehow worth four times that.

Anyone with $40 million worth of something and a shell company can do the same thing.

The bigger issue for me is these business models prey on the poor and overworked. There are a lot of people who can’t buy a $200 backpack because they need that money for food, but they still need a backpack, only have time to grab something at Wal-Mart, and they get stuck in poverty product churn.

It’s like regressive taxation but carried out by capitalists.

This is part of it for sure. It is also true that many open source business depended on it not being worth the trouble to figure out the hosting setup, ops etc, and the code. Typical open source businesses also make a practice of running a few features back on the public repo.

Now I can take an open source repo and just add the missing features, fix the bugs, deploy in a few hours. The value of integration and bug-fixing when the code is available is now a single capable dev for a few hours, instead of an internal team. The calculus is completely different.

You’re fractionally right with GPUs but RAM and SSDs run on different processes at different fabs.

Ding ding ding, and this is why you are hearing about it. It is marketing for enterprise to pay a premium for the next model, with maybe a wakeup call to enforcement agencies as well (which is also marketing).

Codegen for many companies is much less continuous. Security is always on, and always a motivator.

The past few generations I found I was not pleased with their performance, so now I take them weekly to the macbook sharpener at the saturday market.

At what point does active suspension become robotics? Because there are a lot of automakers who should be repositioning to get their multiples up.

And it is extremely likely Nio made that deal to see whether they could embrace, extend, extinguish them.

Right but wasn’t high effort the default effort before? So ultrathink is gone in all but name.

They don’t want to officially disclose the reality because while some users will understand the realities of protecting a product while innovating, many will just realize it means one can go looking for claude 4.5 performance elsewhere.

I’m going in circles. Let me take a step back and try something completely different. The answer is a clean refactor.

Wait, the simplest fix is the same hack I tried 45 minutes ago but in a different context. Let me just try that.

Wait,

If this were true and you were their manager you would not have to deal with it. You could coach or remove them.

The job of execs/middle managers seems to often be dual parenting: 1) coordinate the capable well-parented employees below them, and 2) pander to the usefully myopic spoiled brats above.

Yes exactly, your conception of write includes their conception of erase.

Bennett’s reversible computing is essentially just the way to avoid erasure: whenever you write, keep all the previous states around.

I view people like him like large satellites or small planets that have been crashing through everything mindlessly and creating this wake of wreckage. The only way to keep going is to not look back, or in this framing, not look in.

In single-minded pursuit of a simple goal and with early success they reduce their own humanity so that their repeated actions can maintain their simple function.

Looking anywhere behind/within has become so overwhelming and so painful they will construct elaborate narrative and even engage in medical assistance (eg ketamine) to avoid the consequences of integration.

This is an overstatement of the protection that blanks provide. As it says, they only (potentially) provide insight into contamination caused during the extraction process.

Yeah it’s called a regex. With a lot of human assistance it can do less but fits in smaller spaces and doesn’t break down.

Is this true for anything beyond the simplest LLM architectures? It seems like as soon as you introduce something like CoT this is no longer the case, at least in terms of mechanism, if not outcome.

Those products aren’t typically described as having been “hyped” though — just successful or viral. Hyped has a sort of derogatory/schadenfreude subtext.

Erasure is logically irreversible, writing a bit is not. When you erase a bit you compress the logical phase space of the closed system, which means the missing information has to go somewhere — in this case a couple of very low energy phonons into the larger environment.

That was the limit of just one experimental approach that was peer reviewed and published in a major journal. As you can see there are many experiments validating the limit and none invalidating it.

The reality is that the Landauer limit is vanishingly small. I would encourage you to review the experiment methodology and see if you can come up with better, fundable methods.

This was published right before people started experimentally validating the Landauer limit. I am not sure why it hasn’t been taken down at some point as the evidence has accumulated:

2012 — Bérut et al. (Nature) — They used a single colloidal silica bead (2 μm) trapped in a double-well potential created by a focused laser. By modulating the potential to erase the bit, they showed that mean dissipated heat saturates at the Landauer bound (k_B T ln 2) in the limit of long erasure cycles.

https://www.physics.rutgers.edu/~morozov/677_f2017/Physics_6...

2014 — Jun et al. (PRL) — A higher-precision follow-up using 200 nm fluorescent particles in an electrokinetic feedback trap. Same basic physics, tighter error bars.

https://pmc.ncbi.nlm.nih.gov/articles/PMC4795654/

2016 — Hong et al. (Science Advances) — First test on actual digital memory hardware. Used arrays of sub-100 nm single-domain Permalloy nanomagnets and measured energy dissipation during adiabatic bit erasure using magneto-optic Kerr effect magnetometry. The measured dissipation was consistent with the Landauer limit within 2 standard deviations using the actual the basis of magnetic storage.

https://www.science.org/doi/10.1126/sciadv.1501492

2018 — Guadenzi et al. (Nature Physics) — Opens with:

The erasure of a bit of information is an irreversible operation whose minimal entropy production of kB ln 2 is set by the Landauer limit1. This limit has been verified in a variety of classical systems, including particles in traps2,3 and nanomagnets4. Here, we extend it to the quantum realm by using a crystal of molecular nanomagnets as a quantum spin memory and showing that its erasure is still governed by the Landauer principle.

https://www.nature.com/articles/s41567-018-0070-7

The Landauer limit is not conjecture.

For people who want to ask a model for an app, or a website, or something at a level of “hey you make apps right, I have had this idea for years…” the experience is akin to a slot machine — sometimes they get what they imagined their description would create and it works, and sometimes they get a hollow chocolate approximation.

Only if the model is actually a human or equivalent, otherwise we don’t know what it is.

That’s because it is literally just a feedback loop?

The readme opens with this:

> I have an RTX 5070 with 12 GB VRAM and I wanted to run glm-4.7-flash:q8_0, which is a 31.8 GB model. The standard options are:

> Offload layers to CPU — works, but drops token/s by 5–10× because CPU RAM has no CUDA coherence. You end up waiting. Use a smaller quantization — you lose quality. At q4_0 the model is noticeably worse on reasoning tasks.

> Buy a bigger GPU — not realistic for consumer hardware. A 48 GB card costs more than a complete workstation.

> None of those felt right, so I built an alternative: route the overflow memory to DDR4 via DMA-BUF, which gives the GPU direct access to system RAM over PCIe 4.0 without a CPU copy involved.

And then limps home with this caveat on the closest thing to a benchmark:

> The PCIe 4.0 link (~32 GB/s) is the bottleneck when the model overflows VRAM. The best strategy is to shrink the model until it fits — either with EXL3 quantization or ModelOpt PTQ — and use GreenBoost's DDR4 pool for KV cache only.

I think the reason it refers it to DDR4 is because that is how the user explained it to their coding agent. LLMs are great at perpetuating unnecessary specificity.

It’s funny and also disadvantages everyone who can’t trade 24/7. Win/win?

I’m confused by this discussion. It seems like you said the biases were structural because we know who reports early and that is why the early numbers are always revised down. Structural implies known in advance.

It also seems like you said they shouldn’t revise the numbers but now you are saying they already do.

What am I misunderstanding?

You use a trailing stop loss. You get closed out 15% down from the top, not 15% down from purchase. The alternative in a 24 hour market is worse — the news of a real event hits and by the time you wake up and respond you’re down 50% or more and the stock isn’t coming back.

This policy change is to hunt profit from a safety mechanism used by retail traders.

It is something that should yield a lot of profit for 24 hour trading systems during a downturn.