pixelsort

If this ever became widespread, people would game the "verified real" checkmark using the analog gap.

1. Print AI photo 2. Point fancy expensive camera at printed photo 3. Take a "real" photo

Then you'd see more sensors and even more expensive cameras. Then you'd see miniatured virtual-production volumes.

So, this is not a solution. It's just an arms race disguised as one.

The OP rediscovered that frontier models prefer to reason over logical scaffolding for complex tasks. They excel at technical work with many constraints as they can perfectly maintain the references, flow graph, and evidence states while they works through your conformance gates.

Many people here might disagree; but they're holding it wrong. Those folks should ask:

1. Am I using free tier tokens? 2. Am I working on trivial software? 3. Am I expecting models to adapt to my ways of thinking?

Anyone affect by any of these three mistakes will maintain an impenetrable filter of perpetual ignorance about model capabilities. Since the OP came with receipts, I'm reproducing an example graph below.

From a plan in my active research project:

  Phase 0   resolve -> vendor -> pin             [USER+agent]  -> G0 CORPUS-RESOLVED   [PASS]

  Track B   B1 Abusalah              [WRITTEN]   [agent]       -> G1B Q-BOUNDARY
            B2 Guan-Riazanov-Yuan    [WRITTEN]                  (KILL-BOUNDARY risk)
            B3 Mahmoody-Smith-Wu     [WRITTEN]  + O-MSW-VERIFIER-COST closed (t = 588, MEASURED)
                                     -> G1B PASSED: BOUNDARY-ESTABLISHED

  Track A   05 Binius                [WRITTEN]   [agent]       -> G1 AUDITS-COMPLETE [PASS] 2026-08-28
            07 FRI-Binius 2024/504   [WRITTEN] — vendored by USER; calculated Q-FLOOR point obtained
            06 HOBBIT                [WRITTEN] — O-QUEUE-M escalated to 179 Phase 2
            01 Wesolowski            [WRITTEN] — hinge splits; C_overlap = 0, C_tail = 57.50% computed
            02 Tight VDFs            [WRITTEN] — no omega; O(log T) proof threads; black-box RO base impossible
            03 LaBRADOR (light)      [WRITTEN] — calculated, not measured; translation remains open
            04 LatticeFold+          -- deferred to pre-141

  Track C   [PASS] C0a contract -> C0b exact simulator -> G0C TRANSLATION-READY
            C2 Alwen-Serbinenko [TECHNIQUE-ONLY] -> C1 SoW [PASS] ANCESTRY-FOUND
                                                -> C4 Seeded PoW [FORM-ONLY]
                                                                 -> G1C [RED] WORK-LEMMA-MISSING
                                                                          [WARN] ANCESTRY-CLASSICAL
                                                                 -> L0 [PASS] SCOPE-FROZEN
                                                                      |- E0 response [PASS] PRESENT
                                                                      `- L1-L2 local [PASS] CARRIER-QUOTIENT
                                                                           -> J0 [PASS] TECHNIQUE-FIT CONFIRMED
                                                                                -> admission decision [PASS] CANONICAL RESULT
                                                                                     -> A0 [AMBER] REFERENCE-TESTED
                                                                                          -> A1 [PASS] ROOT-LEAF FEASIBILITY LEAD
                                                                                               -> semantic equivalence + epoch [USER]
            C3, C5  -- deferred to pre-141

  Phase 2   Q-TAIL / Q-FLOOR synthesis           [agent]       -> [PASS] G2: TAIL-MECHANISM-FOUND · FLOOR-OPEN
  Phase 3   reference ledger for 141              [agent]       -> CLOSED; awaits fresh USER authorization

AI will not begin to benefit student learning unless educators pivot into a strict bifurcation pattern around access.

For example, alternating weeks with and without the permission to use AI assistance. This is also the latent solution for corporate de-skilling that no one seems to be talking about yet.

However, it won't happen (in both cases) because we're too early. Domain experts like the OP are still relinquishing their expertise to unlock productivity gains. Once that bottoms out there will be no longer be demand for people like OP to betray their peer groups.

Learning with AI assistance be as frictionless and productive as I suppose it can, but de-skilling and job consolidation (role expansion) will continue.

I've been using my own variant docset driven development since late last year. It emerged organically for me by adapting to the limits of the agents.

- Prefix-numbered sets of documents, keeping them clustered visually and easily referenced by humans and agents. (124.5 = docset 125, phase 5) - Each doc gets a suffix. Most of the time STRATEGY is the first doc. Sometimes NOTES, DISCUSSION, or HANDOFF though. - Once the strategy doc is comprehensive, the multi-phase checklist PLAN doc can emerge. - During execution, the agent drives the PLAN while purposefully expanding lightly scaffolded phases before entry. - As the user, I drive a PROGRESS doc during the PLAN execution. This tracks the progression of my own questions and the important work summaries that I need for tracking the current work trajectory and for historic analysis over past decision-making. - Any concern that exceed the scope of the active docset numbers calls for a new one. - A "docs/archive" folder is kept where I sweep docsets into numbered eras.

I have aliases and templates too, but DDD is so simple that it's overkill.

Any modern reasoning agent can take the plain explanation above and understand you effortlessly when you say "open a new docset strategy" or "read docset 125 and proceed".

VS Code does actually already have this. I use it in Antigravity IDE.

https://marketplace.visualstudio.com/items?itemName=anthropi...

I'm working on L1 blockchain that turns your hardware into a productive mining asset by leveraging my Rule 30 VDF construction as addressable but irreducible sequencer. The ZK proof system is recursive FRI-Binius and utilizes a binary field tower.

Yesterday, I finally achieved blockchain payload byte-count stability from block 4 and onwards. Today I'm pulling levers to reduce the payload size. Proof-of-everything, so no confirmations needed.

https://bitcointalk.org/index.php?topic=5580132.new

Yes, that's probably his dumbest public idea to date. Given that this GPT repos and parts of autoresearch are brilliant I'm sympathetic. I think he's earned the right to exhibit mild expressions of AI psychosis at this point.

And, my objection was that he clearly had no understanding of the supply-chain risk he was worsening by advocating widespread use of Obsidian for agentic engineering tasks.

Since his announcement, Obsidian has taken proactive steps to mitigate the risks, or at least study threat model. Hopefully, they will implement proper RBAC or something before someone else with his visibility announces an even more irresponsible half-baked idea.

Make sure you are running adaptive cooling (or just bump them up) on the top fans in your case. Also, ensure that you are undervolting appropriately using something like MSI Afterburner or GreenWithEnvy.

If you don't, you could easily toast your RAM -- especially under BF16.

There could be opportunities we haven't anticipated.

What if labor organizes around human work and consumers are willing to pay the premium?

At that point, it's an arms race against the SotA models in order to deepen the resolution and harden the security mechanisms for capturing the human-affirming signals produced during work. Also, lowering the friction around verification.

In that timeline, workers would have to wear devices to monitor their GSR and record themselves on video to track their PPG. Inconvenient, and ultimately probably doomed, but it could extend or renew the horizon for certain kinds of knowledge work.

In the many darker timelines that one can extrapolate, capturing essential tech stacks is just a pre-cursor to capturing hiring.

Once we start seeing Open AI and Anthropic getting into the certifications and testing they'll quickly become the gold standard. They won't even need to actually test anyone. People will simply consent to having their chat interactions analyzed.

The models collect more information about us than we could ever imagine because definitionally, those features are unknown unknowns for humans. For ML, the gaps in our thinking carry far richer information about is than our actual vocabularies, topics of interest, or stylometric idiosyncrasies.

The level of compliance and enthusiasm varies. Some believe they are making the world a better place. Some feel they're adding value but suspect they are trapped within a cycle they refuse to examine. Some are more connected to the truth, and comply willingly but resentfully.

Where you fall depends on where you work and what you work on.

You make a great points about the chain of accountability. But, in my opinion, working professionals are the only agents in the system with the potential to realize their own culpability and divert their actions.

Perhaps, it isn't fair to point to them and call them traitors. Still, they are the only ones with enough agency to potentially organize and collectively push for the kind of ethics that could save us all.

While the desire is not new, advancements in LLMs and diffusion models have made this sort of bridging effective and attractive to an unprecedented degree.

Those massively and widely available benefits will continue to deflate the value of human intelligence until even most of innovators currently working on them lose their seats at the table too.

> There is every reason to believe that those who invest in deep understanding will continue to be valuable, regardless of what tools emerge.

I don't take issue with this, except that it's a false comfort when when you consider the demand will naturally ebb and individual workload will naturally escalate. In that light, I find it downright dishonest because the rewards for attaining deep knowledge will continue to evaporate; necessitating AI-assistance.

The reason is it different this time around is because the capabilities of LLMs have incentivized the professional class to betray the institutions that enabled their specializations. I am talking about the amazing minds at Adobe, Figma, and the FAANGS who are bridging agentic reasoners and diffusion models with domain-specific needs of their respective professional users.

Humans are class of beings, and the humans accelerating the advance of AI in creative tools are the reason that things are different this time. We have class traitors among us this time, and they're "just doing their jobs". For most, willful disbelief isn't even a factor. They think they're helping while each PR just brings them closer to unemployment.

Re: Permanence

I mean "permanence" in the same vague senses that I think the OP was hinting upon. A belief that regardless of change, the primitives remain. This is about having total confidence that abstractions haven't removed you the light-cone of comprehension.

Re: Appliance

I believe turing-completeness is over-powered, and the reason that AGI/ASI is a threat at all. My hypothesis is that we can build a machine that delivers most of the same experiences as existing software can. By constraint, some tasks would impossible and others just too hard to scale. By analogy, even a Swiss-army knife is like an appliance in that it only has a limited number of potential uses.

Re: Users

The machine I'm proposing is basically just eBPF for rich applications. It will have relevance for medical, aviation, and AI research. I don't suppose that end-users won't be looking for it until the bad times really start ramping up. But, I suppose we'll need to port Doom over to it before we can know for sure.

My goal isn't to make LLM-assistance impossible; it will still be possible. In fact, GPT2-level inference is one of launch demos I have planned if I can finish this cursed self-hosting run.

My goal is to make training (especially self-training) impossible; while making inference deterministic by design and highly interpretable.

The idea is to build a sanctuary substrate where humans are the only beneficiaries of all possible technical advancements.

I was 7 in 1987, learned LOGO and C64 BASIC that year, and I relate to this article as well.

It feels as though a window is closing upon the feeling that software can be a powerful voice for the true needs of humanity. Those of us who can sense the deepest problems and implications well in advance are already rare. We are no more immune to the atrophy of forgetting than anyone.

But there is a third option beyond embrace or self-extinguish. The author even uses the word, implying that consumers wanted computers to be nothing more than an appliance.

The third option is to follow in the steps of fiction, the Butlerians of Dune, to transform general computation into bounded execution. We can go back to the metal and create a new kind of computer; one that does have a kind of permanence.

From that foundation, we can build a new kind of software, one that forces users to treat the machine as appliance.

It has never been done. Maybe it won't even work. But, I need to know. It feels meaningful and it has me writing my first compiler after 39 years of software development. It feels like fighting back.

Thanks. The hardest part has been slogging through the segfaults and documenting all the unprincipled things I've had to add. Post-bootstrap, I have to undo it all because my IR is a semantically rich JSON format that is turing-incomplete by design. I'm building a substrate for rich applications over bounded computation, like eBPF but for applications and inference.

Currently striving towards my own TypeScript to native x86_64 physical compiler quine bootstrapped off of TCC and QuickJS. Bytecode and AST are there!

I never appreciated tailwind until AI models revealed it as such a token-efficient way transport styles between models and other use-cases. AI aruably hurts demand for their premium offering the same way it hurts demand for junior devs.

I'll sometimes ask Claude Sonnet 4.5 for JS and TS library recommendations. Not for "latest" or "most popular". For this case, it seems to love recommending promising-looking code from repos released two months ago with like 63 stars.

Don't forget that OpenAI was also following Anthropic's lead at the model level with o1. They may have been first with single-shot CoT and native tokens, but advancements from the product side matter, and OpenAI has not been as original there some would like to believe.

Seems like the opinion of someone who doesn't know that OpenAI cloned Anthropic's innovations of artifacts and computer use with their "canvas" and "operator".

I suppose it is time to finally apply to YC for DeepMojo. With over 40 packages in our monorepo, after a year of ramping up, we recently launched on Windows, and achieved MacOS support internally. Secure, local first, zero-trust AI, with opted-out defaults, no user telemetry, no ads, and no internet required.

Fitting then, that his release is happening to make a point with respect to judicial overreach in New York.

Also his opsec was sloppy. If you want to believe that the spooks were doing full ipv4 scans to DDoS all his legit exit nodes that would make a better movie. But really, he was just in over his head.

Predictably, dark web market operators adapted afterward. The state got lucky and they knew it, so that also factored in to their sentencing recommendations.

Glad he's getting out.

Does "suddenly politically inexpedient performative pandering program he calls a fact-checking" solution work better? I didn't want to go in for the edit.

A tasteful post and distinction well highlighted. Humorously, Yvo Schaap is no stranger to 10x thinking. For one thing, Yvo publishes diagrams on SaaS/dev topics that always seem consistently way ahead of their time in terms of their organization and completeness.

Roots? FB's roots are frat boy pranks and backstabbing your actual friends. Better headline: Billionaire backtracks on freespeech after a private meeting with a much more powerful billionaire to discuss ways of making amends for his pesky commitment to a well-informed society.

I imagine that with native tokens for planning and reflection empowering the models I'm referring to, it is something like a search space where we've enabled new reasoning capabilities by allowing multiple progressions of gradient descent that leverage partial success in ways that weren't previously possible. Lipstick or not, this is a new pig.

What has changed with CoT and high compute is not yet clear. My point is that if it makes bare prompt injection harder for humans then we shouldn't call it a fundamental limitation anymore.

Are LLMs nothing more than auto-regressive stochastic parrots? Perhaps not anymore, depending on test time, native specialty tokens, etc.