The OP specifically mentioned coding in regulated environments, so "needles" (specific pieces of information) are important and this does degrade with context size.

https://benchlm.ai/benchmarks/...

You also take a price hit if your context cache times out (assuming your use a provider) and that adds up.

I was responding in the context of the OP. You absolutely can use Qwen3.5/3.8 locally to good effect (and I use abliterated 3.5 locally). What I was talking about is people claiming their laptops blast out 100 toks/sec with an agent swarm. That's the horseshit.

If you try to one shot it, you're going to have a bad time... even with the best frontier models.

- Have clear, specific requirements (with no conflicts/confounds)
- Use a more expensive frontier model to plan the workflow in discreet units (context rot). You should target using at most half the context for a session without having to compact.
- Use capable models to handle the discreet units with TDD (test driven development)
- Do adverserial reviews with a different model family that has low sycophancy (tendency to agree with the user)
- Have checkpoints where you review decisions and judge the results

- If you're using openrouter don't let it float on any provider. Use presets to filter out heavily quantized models (fp4, etc) if you start getting really squirrelly results, in the middle of an otherwise good run, look at the openrouter logs. You probably failed over to a bad provider. Filter them out with the preset or guardrails. The prompts should specifically instruct it to come back to you if it has any questions. Newer models (I'm looking at you Opus) will try to figure it out for you without saying anything and this is where things often go sideways

This will usually get you 90-95% there, depends on the language and the project. If it's something without a good training sets and documentation, you're still going to have a lot of issues (e.g. we had a Mulesoft project where I ended up hand coding a lot of things because the training set is weak and Mulesoft's documentation is horrible... the models just didn't have enough information).

I use this for larger projects and the are always minor issues but the results, especially with the latest generation of models, the overall output is strong and saves a lot of time.

Finally, don't believe the crap you see on SNS. Unless you have some really expensive hardware at home... you're not going to be able to do anything useful with a local model and you're definitely not going to be able to do it fast. Anyone who tells you that produced anything serious with a local Qwen with q4 weights is lying or doesn't make anything really functional.

Not a single one of your articles supports your conclusion. They're actually arguing for better (not more, better) management:

- The forbes article is an opinion piece stating that remote work exposes weak managers. It argues that WFH needs _better_ managers.
- LSE argues that well-trained managers _boost_ productivity for hybrid/WFH.
- National Law Review is also arguing for better trained managers (specifically communications) because they also make hybrid/WFH more effective.

All them state what has always been the case, effective managers who communicate tasks and objectives well and then step back and let employees actually do their jobs improve things. Micro-managing/toxic bosses are still counter productive regardless of the environment.

You'll argue "If I'm doing my job, I don't need managers." A good manager should be taking care of the sh*t you don't need or want to be doing. Either doing it themselves or seeing that someone else is taking care of it and not blocking you.

The current consensus for peer-reviewed research is that the best, most productive setup for most companies is actually hybrid ... targeting 2 days in the office. It depends on the industry, work being done, etc. RTO and full WFH are generally not the best setup.

So far most of the popular Chinese models have relied heavily on training with US Frontier models. Even as recently as Qwen 3.5 it would sometimes respond that it was ChatGPT or Gemini. Anthropic and ChatGPT even play whack a mole with accounts that are doing this and a large chunk of them have been tied to the PRC. It's a perfectly valid approach (technically) but it's against the ToS for the providers.

The slides that are circulating say the quiet part out loud. Itâ(TM)s one workload (timed compilation) and they position their 20 core part as âoebeatingâ a 10 core M5 and coming âoereally close guys!â to a 16 core M5 pro. Whatâ(TM)s missing? Any LLM benchmarks⦠itâ(TM)s a repeat of the DGX which was quickly repositioned as âoemeant for low cost development for the CUDA stackâ after the M4s inference mopped the floor with it. (MLX/metal does have limits but for general purpose use itâ(TM)s not even a contest).

Neither of you are actually citing any information to back up your positions. Both arguments boil down to "nuh-uh, you're dumb." For instance:

- Total undergraduate placement (workforce, post-grad, fellowships) over the past 20 years has held steady ~85%

This is just a very shallow pass but it would seem to support Local ID10T's position that DEI has no impact. A much more through review would be required. However, if you go back 50 years you notice something different. The undergraduate workforce placement rate was 25%. It's now at about 60% (peaking at 70%). Before the 80's, Harvard was seen as a stepping stone to a research and academic career... not "leaders of industry".

Anyway, you can continue pointless name calling... or you can have a discussion. Your call. One often fixes things, the other generally doesn't.

It's an 16x RTX Pro 6000s.
- 1.6 TB/s GDDR7 vs 8 TB/s HBM3e on a B200
- No NVLink, just PCIe Gen5 which is roughly two orders of magnitude slower
- A gigabit switch... _gigabit_

That's not a frontier model server, that's an AI workstation. And no, I'm not exaggerating. You don't run models production models over PCIe bridges.

And, if you continue reading the docs you'll see that the XFRA model has account data, chat history, system prompts, all the supporting infra back at the data center. So you have a substantially slower inference that still has to phone home over a 1 gigabit _symetric_ link for the data it needs.

And that's on top of the fact this "distributed" data center thing has been tried before and doesn't work for many practical reasons.

This has to be an AI bubble grift. I can't think of a single sysadmin I know that wouldn't kick you square in the nuts for even suggesting that you do this with their infrastructure.
- The last mile power and comms is questionable (the grid may have 40% capacity)
- The availability will suck
- The latency will suck
- The downtime until a operator can get out to fix it will really suck
- Homeowners will kick them out after the first 3:00am service call

The industry is not going to die. Players will but the industry wonâ(TM)t. AI is overhyped but it is very real and functional. The hype is causing companies to get over extended and when the investments dry up, theyâ(TM)re screwed (Iâ(TM)m looking at you OpenAI).

Google is just hedging and also trying to lock in Anthropocene as a Cloud customer.

I'll preface this by saying I don't think LLM creators should be able to use content without permission/license. This is just an interesting discussion.

LLMs generally do not reproduce text. They can be made to do so with specifically crafted prompts but no current LLM is just going to regurgitate "Free as in Freedom" unless asked to do so. Instead it will use statistical matching to apply the text to probable matches, a very crude version of what we do. LLMs are starting to approach the way we meat sacks use books. We take in the information and then we apply it to problems. Where do we cross the line? Where do we say anything (or anyone) who is trained on (has read) this material is now required to do their work for free because they have the knowledge from that book as part of their training set?

It seems a little preposterous but that's where this is headed logically. It's shifting from "You can't reproduce this book." to closer to "You can't use the knowledge in this book except under the conditions we dictate." That's dangerous.

The details are freely available in the bottom of a locked filing cabinet stuck in a disused lavatory with a sign on the door saying ‘Beware of the Leopard".

Dear young turks:

If you can't take it with you when you decide to walk out the door (or the company shows it to you), it is not compensation. Period, end of story.
- Stock options: do you actually have the stocks yet?
- Bonuses: do you actually have the bonus yet?
- Fancy coffee machines and free lunches: nope
Insist on money so that you can decide what to do with it.

A note on stock option. Besides the fact that stock options often don't pan out, investing in the industry you work in (much less the company you work for) is generally a bad idea. If the industry is in a down cycle and you get laid off, your options are probably worthless at that point as well. The value of your options is not tied to your effort but the decisions of PHB a few floors up.

Sincerely,
The Grey Beard

- The amount of energy and effort to deliver that 15 watts (food) is around 150W (10:1) based on the US needs.
- It takes years of investment before that one human brain can function at that level.
- There's also a decline at the end where it's not delivering what it used too.
- There are also all the additional creature comforts that we need to function (lighting, air condition, etc, etc.)

With that said... AI's are still no where near the same value proposition as the human brain (I actively use them and push for their _effective_ use at my job), much less providing actual companionship and "human" value to life. Him making this comparison is ridiculous.

Whatever the case, Thesis #2 has got to be a part of it. Cloudflare and Akami are two of the stocks taking a hit and that makes no sense. I don't care how good the code is:
1. I want someone else to handle the brunt of the DDoS crap so I can focus on running my app.
2. AI doesn't magic a global CDN for you.
People saw AI + Security and just started dumping anything related to security without understanding what Anthropic's announcement was really about or what the companies they were dumping really do.