adam_arthur

I'm talking about automation generally, not agent loops.

E.g. prompt A to achieve X, output in format Y. Use Y to do something in prompt B.

Agentic loops will underperform deterministic control flow pipelines (with non-determinism constrained to LLM calls).

Agents are more general, which is the main advantage. But inherently a more general solution will waste context on unnecessary reasoning.

Try asking the smaller Qwen models to output a JSON in a specific format. It basically can't do it consistently with a moderately sized prompt unless you constrain the token generation via GGML or are extremely repetitive and specific about it. (Thinking disabled)

Gemma 4 will do it correctly pretty much 100% of the time. (Thinking disabled)

Applies to other rule following as well in my experience.

Qwen may be better at toolcalling and certainly probably codegen.

It seems to me Google explicitly designed Gemma for edge device automation, and didn't fine tune for agentic or coding use cases.

Yes, I'd recommend a 5090 over the DGX Spark if your goal is general automation.

You can run multiple instances of these models in parallel on the DGX Spark which somewhat mitigates the difference if your task is parallelizable.

But I'd take the simplicity of a single thread and higher throughput personally.

Overall of course still better to wait for next gen devices if you can.

Gemma 4 is particularly good at pipeline/automation tasks.

It outperforms all the Qwen models (even 100B+) for rule following/automation style tasks in my experience. Its image interpretation is also very good, and out-benchmarks Opus.

Qwen seems to ignore instructions and consistently outputs incorrect formats (when token generation format is not explicitly constrained)

But yes, on the DGX Spark Gemma 31B Q4 with MTP runs around 20 tok/s and Gemma 26B A4B around 60 tok/s. Still quite slow. But on a high end Nvidia card would run significantly faster and still fit in memory.

I'd recommend for anyone getting into local models to focus on memory bandwidth over RAM. Models under 100B parameters are now sufficient and hugely useful for automation.

I agree that for coding/creation use cases, there's still not a compelling argument for local models.

But e.g. if you want to scan a list of stocks and interpret news/high pass filtering, interpreting logs, interpreting screenshots, the local models are more than sufficient already.

The lower the code quality and less coherent the design, the poorer the agent will perform over time as the project grows.

Pure vibe coding pretty much limits the max scope of your project to the capability of the agent, and if you're trying to make something real and maintainable, you'll always end up with something worse than the person who guides it at a lower level.

That being said, there are certainly going to be systemic/built in ways to have coding agents generate coherent architectures. Just it isn't really baked in right now.

Probably an ideal model with current tech is the main coding agent thread that implements things, and a secondary thread that constantly compacts/prunes and cleans up according to specific style rules.

Problem with doing both in same context via AGENTS.md is the LLM often won't adhere to all rules if it's overloaded with many competing instructions. But you can scope the code styling agent prompt tightly and get much higher focus and adherence.

There is no way to prove, but what is definitely true is that many people are attempting to use LLMs on forums and otherwise.

So if you think none of these comments are written by LLMs, you're probably mistaken too.

In the end we accept that we can't tell anymore and move on (barring some biometric protocol that can't be gamed via automation)

There are many with subtle tells.

Not nearly as obvious as the ones from 6 months ago, but seems to be more the use of hyperbolic phrasing in a particularly unnatural way.

The assess/explain, then hyperbole at the end kind of structure.

Top comment looks suspicious from this perspective, but it's kind of a losing battle to be able to differentiate them with sufficient accuracy anyway

Anthropic is absolutely not profitable on a GAAP basis.

Companies may state cash flow positive, operating profit, EBITDA positive, but this is not a true profit in aggregate, just when stripping out many other expenses.

If anyone has evidence to the contrary, please share. Once they go public it will all be free to review, at least.

There's quite a large cushion for software salaries to decline before permanent structural unemployment were to set in.

It's not really feasible for "normal" businesses to hire developers at current salaries.

Tech companies will probably shrink in headcount, but all the non-tech kind of businesses can increase developer headcount.

Current Tech salaries are far above other fields while requiring (used to) significantly less training or time investment to get into.

Phase 1 is more likely that software comp will normalize with other professions, and more hiring will happen at the fringes rather than being concentrated in a few big companies.

These small models exist in the cloud and are/will be priced commensurately to their size.

Your claim is effectively that companies don't care about operational/cloud costs. Even pre-LLM, companies regularly assessed and tried to pare down cloud spend.

The entire universe of automation projects that can be run effectively for free relative to SoTA models?

I don't think many realize that most LLM embedded automation, pipelines, products will soon be able to run extremely cheaply on models < 100B parameters.

Frontier models will be used for coding/creation use cases, yes. But for all the pseudo-deterministic, pipeline, analysis style things there will be no practical benefit to running frontier models, only additional cost.

Gemma 4 26B outperforms most 100-200B models that I've tested for reasoning and structured output.

Gemma 4 12B can consistently select where to click on browser images given a minimal prompt, and do so very quickly.

If you look at all the tech companies doing buybacks, usually the shares created for employee RSUs matches or exceeds the shares retired from buybacks.

Not in all cases, but many

Which is why GAAP earnings matter and not free cash flow

Pretty much everything that appears novel in life is derivative of other works or concepts.

You can watch a rock roll down a hill and derive the concept for the wheel.

Seems pretty self evident to me

I had the complete opposite experience.

The more money and security I made, the less beholden I felt to the workplace and things that weren't personal to me. Less anxiety, less need to appeal to others for survival/success (of course still try to be nice)

I've also always had tons of ideas and things I wanted to pursue outside of work. So never lacked for motivation or sense of purpose.

My dad recently retired and seems to not have much direction now, so it must be a somewhat individualistic response.

Always interesting to see others reactions!

In the rare case that all your concepts use the exact same descriptive word, you are probably right!

The majority of the time you can infer the type from reading well written code (to the extent that the shape of the type matters in the context of that piece of code)

It seems clear to me from first principles.

Humans are trained on human language. LLMs are trained on human language.

Thus something that is easier for a human to understand is likely easier for an LLM to understand.

That higher level language with well named variables reads more comprehensibly than code:VERB with:PREPOSITION types:NOUN, intermixed:ADJECTIVE, stems:VERB from:PREPOSITION first:ADJECTIVE principles:NOUN too:ADVERB

Strong typing has clearly won.

However, verbose typing is likely a negative for LLMs.

Algorithms written in "pseudo-code", aka a higher level language without type information, are far more readable to a human, and thus likely an LLM too.

In regards to control flow and general concept of what code is doing, types provide very little info over well named variables. In fact they often impair understanding by breaking up logic with implementation details.

I'd be curious to see some experiments around this, but I'd guess strongly typed languages where the type information is mostly hidden/inferred would have better generation accuracy from a semantics perspective (and likely worse from a type safety perspective, but can be corrected on compile/retry)

I recently found Gemma 4 e4b surprisingly effective for small "classification" style tasks for something I'm doing at work.

In this case, picking out "semantic" css classes on single dom nodes.

Was able to run it on my 4(?) year old M2 mbp with 16GB of ram and it runs in only 100ms or so per query. Probably it can run much faster, but haven't experimented with batching etc

With tight and targeted context control, you can use extremely small models for useful things. Ideally with problems where the harness can be mostly deterministic and you have known bounds on what you're trying to do

The blackout started back in January before the US even got involved.

Due to widespread protests and an attempt to crack down on coordination. This chain of events was widely reported.

https://en.wikipedia.org/wiki/2026_Internet_blackout_in_Iran

Fear can create a self fulfilling problem, even if the fundamentals don't warrant that problem.

I don't know in this case, but I'd guess

1) Cyber security risk (straightforward)

2) Software stocks have been declining materially on AI competition fears, which can potentially impair loans to those companies, which cascades if there are enough defaults.

The second is mostly just driven by misinformed investors IMO, but the valuation haircuts as a result are real and can impair existing loans.

(Not to say these software companies weren't overvalued to begin with, just the reasoning for the decline is misplaced imo)

On the other hand, I never understood the focus on computer use.

While more general and perhaps the "ideal" end state once models run cheaply enough, you're always going to suffer from much higher latency and reduced cognition performance vs API/programmatically driven workflows. And strictly more expensive for the same result.

Why not update software to use API first workflows instead?

I don't understand the position that learning through inference/example is somehow inferior to a top down/rules based learning.

Humans learn many, and perhaps even the majority, of things through observed examples and inference of the "rules". Not from primers and top down explanation.

E.g. Observing language as a baby. Suddenly you can speak grammatically correctly even if you can't explain the grammar rules.

Or: Observing a game being played to form an understanding of the rules, rather than reading the rulebook

Further: the majority of "novel" insights are simply the combination of existing ideas.

Look at any new invention, music, art etc and you can almost always reasonably explain how the creator reached that endpoint. Even if it is a particularly novel combination of existing concepts.

Which is exactly how humans learn many things too.

E.g. observing a game being played to form an understanding of the rules, rather than reading the rulebook

Or: Observing language as a baby. Suddenly you can speak grammatically correctly even if you can't explain the grammar rules.

The 9% of borrowers defaulting stat cited in the title is not the same as 9% of the loanbook defaulting.

As stated in the article, 9% is the number of borrowers that defaulted, which was concentrated in smaller borrowers (thus smaller loans).

And then, again, you can say probably half of the dollar amount of those defaults are recoverable.

Bond defaults spiked to around 6% in aggregate in 2008, to use a worst case example.

It sounds like you're effectively describing some fraud scheme.

A smart lender will not issue loans without real collateral. If you create a subsidiary, that subsidiary has to have sufficient collateral and cashflow to secure a loan.

> Oooh, source? (I'm curious for when this was measured.)

It depends when you measure, but you can Google around and find figures in the 60-80% range. 80% may have been a bit on the optimistic end of the range. But it's important to note that a "default" doesn't imply a 0.

Of course this will depend on the covenants, underwriting standards, type of collateral.

I would guess software equity collateral recovery rates are lower than hard assets like a building. (Which is why I personally don't like Software loans, nothing to do with AI)

> Correct. The question is if 20% is enough, and if a 20% markdown creates a vicious cycle as funding for e.g. re- or follow-on financing dries up.

I think it's almost certain that new fundraising for private credit will be materially hindered going forward. But this just limits the growth rate of these firms, does not introduce any "collapse" risk.

They may move from net inflows to net outflows and bleed AUM over a period of some years.

If NAVs were inflated previously, they may be forced to mark down the NAV to meet redemptions rather than using inflows to payoff older investors.

In the world of credit, 20% is an enormous haircut. Again, senior secured loans fell by around 30% peak to trough in 2008.

We have the public BDC market as a comparison point where the average price/book is around 0.80x. So the public market is willing to buy credit strategies at a 20% discount to stated NAV.

The real systemic risk here, if we were to reach for one, is really that these fears become self fulfilling.

If investors pull funds out of credit strategies en-masse, there is no first order systemic issue, but it means borrowers of many outstanding loans may not be able to secure refinancing as money is drying up.

This could lead to a self-fulfilling default cycle. But this would be a fear driven default cycle, there is no fundamental issue with cash flows of borrowers or otherwise (in aggregate, currently).

Finally, in regards to the asset managers themselves, many are quite diversified.

Yes, they have private credit funds, but many have real estate funds, buyout funds etc. OWL is one of the biggest managers of data center funds, for example (which they also got hammered for on AI bubble fears)

Given how depressed pricing is in public REITs, for example, I expect a lot of asset managers to pivot towards more real asset funds.

There is so much misinformed fear-mongering about private credit right now.

Important Facts:

1) The majority of private credit funds are classed as "permanent capital". When you put money into these vehicles, you give the Asset Manager discretion over when to give the money back. Redemptions are often gated at ~5% per quarter.

(So there cannot, by definition, be a run on the bank)

2) Credit is senior to equity, so if you expect mass defaults in private credit, it means the majority of private equity is effectively wiped out. Private equity has to be effectively a 0 before private credit takes any losses.

3) The average "recovery rate" for senior secured loans is 80%. Even if private equity gets wiped to 0, the loss that private credit incurs is cushioned significantly by the collateral backing the loan. These are not unsecured loans the borrower can just walk away from.

(The price of senior secured loans dropped by ~30% in 2008, as a worst case datapoint)

4) Default rates on many of the major private credit managers is ~<1% in recent years. There are other estimates stating higher default rates, but that often classifies PIK income as a default. A loan modified and extended with added PIK that ultimately gets repaid is not a "true" default.

5) Finally, it's true that NAVs are likely overstated, but generally it's by a modest amount. Every Asset Manager today could go out tomorrow, mark NAVs down by 20% and suddenly there is no crisis.

(The stocks of Asset Managers have already traded down such that this seems expected and priced in anyway)

No, it's bad policy.

Cliffs in policies will always lead to players working around the cliffs.

E.g. in NYC there is an additional 1% sales tax on home sales above 1 million dollars.

So nobody in the market would ever sell a home between 1m and 1.01m as the tax increase is greater than the sales price.

These are failed policy implementations (in the above example the tax should be marginal, not thresholded)

Any policy which does not account for individual actors optimizing financially is a badly designed policy.

There are numerous similar examples re: CRE when requiring subsidized housing units for certain sizes of development. Often it's more lucrative to build smaller and get around subsidized unit requirements.

You can call them "asshats", but I'd rather live and discuss policy in reality.

Many of these new, clearly strictly punitively intended, taxes aimed at the wealthy will have the same logical outcome.

Show me the incentive and I'll show you the result

I'm not familiar with that specific example, but I do know that independent players in any economic system will follow the incentives.

Expecting companies, people etc to do the "right thing" when it's financially disincentivized usually doesn't work out.

Same will happen in regards to all these new taxes reinforcing existing population migration trends.

The current governor is proposing cutting property taxes in ~half by eliminating the school district portion and instead funding schools directly via the state's budget surplus.

Remains to be seen, as the next legislative session isn't until 2027.

Oil futures (months out) are priced lower than spot, presumably due to anticipation that Iran driven disruption to the market will be short lived.

(Remains to be seen whether that's true)