The goal is in the first paragraph: "aim was to let consumer gear that can't bitstream TrueHD Atmos render real object-based Atmos with height."
Summary is also clear: ffmpeg doesnt have channel coupling and proprietary cryptographic blocks it.
I was once wondering what would be the problem of doing such a project. Now it's clear to me what problems I would run into and I don't have to burn my tokens.
> # Why It Can't Fully Work
> # Wall 2 — EMDF keyed authentication (the decisive one)
> coregraft = a real Dolby core + our metadata → Dolby Surround. The metadata is rejected too, and we traced it to the EMDF emdf_protection field, which is a keyed authentication code, not a computable checksum:
or
> EMDF container (ETSI TS 102 366 Annex H) — wraps OAMD (id 11) + JOC (id 14) with the emdf_protection field, carried in the E-AC-3 audio-block skip field exactly where real Dolby streams put it (recomputing frmsiz + crc2).
These are I think (not being a Dolby or Audio expert) a mix of exactly what I was talking about. Dense / jargony and also weirdly low information at the same time.
What does older mean in this context? Because some people still think the year 1996 wasn't that long ago. Modern Renault cars are fine and reliable enough. I've had 4 in my life time and had zero issues myself. I see a ton of them here in the UK and, again, they're fine.
There were models with tons of problems, other that were bullet proof really.
I think if we take french cars (Renault/Peugeot/Citroen) in general, most major reliability issues have been on diesel cars exhaust gas recirculation systems due to strict european emissions and they are far from the only brands suffering from that.
German cars were known for their great reliability in the early 90's but in later decades had all sort of electronical gremlins.
Also I think regardless of their actual current reliability, some brands or models attract on average different kind of owners which impact how actual services are followed, if the car is stored inside or outside, if the owner take care or not of warming up the engine in the morning or floor it while cold, and the general care they apply to it.
> maintenance problems” pretty much sums up a lot of older Renaults.
It was in the 2000s but not anymore.
If you go to eastern Europe, specially in the Balkans, you will see a lot of taxi drivers with Renault with milages over 500k km. They do hold the space with the usual Toyota Prius.
The current brands in the EU with bad reliability issues are Stellantis with infamous Puretech engine. And BMWs, not so much for the reliability aspect, but due to the stupidly high service costs.
The brand new BMW I last drove has network issues, it's been to the shop 10+ times the infotainment system glitches worse then skrillix and adjusting the side mirrors crashed and the car had to be turned off and on again to regain functionality. My friend is pulling teeth to get a new replacement, I had nightmares that it would glitch and emergency break us to death on the highway, not confidence inspiring!
show me a prompt that is meaningfully expertly crafted beyond just providing Do's, Do not's, task context, and a goal.
> Correctly prompt, to steer it, to verify it, and to improve the harness.
I doubt this a lot. The average AI user is running claude code as the harness, or Codex etc. prompting has no secret incantations, and steer and verify is just knowing what the answer should roughly look like, which is a domain skill, not an AI skill.
I feel like you don't have any friends who make software but don't know how to code.
Yes, they do make software now - whereas it was impossible before. You may be absolutely shocked at how bad LLM code can be when prompted from a noncoder. How buggy, and how absolutely rife with security problems it can have. I honestly don't know how they can get LLMs to write such bad software - but somehow they can. This is from people who have been vibe coding for 3 years straight btw (huge amount of time p/day).
It would be like Samsung: Samsung does semiconductors, smartphones, displays, TVs, appliances, batteries, shipbuilding, construction, insurance, hotels, and even theme parks.
Actually no. Asians have common decency to run separate businesses as separate companies. Samsung Electronics does a lot. But it really doesn't have direct hand in other business. Even if there is some ownership in them.
For Switzerland it would be significantly worse since they are surrounded by Germany, France, Italy, and Austria. So all trucks would be stopped at border. Food prices are already super expensive and this would make it worse.
They think it's a failure because their political class refused to use the new powers to implement the will of the people on immigration, not due to economics (which, again, were not affected).
Brexit was a referendum on immigration. It was the number one issue. Yet, successive governments since Brexit have turned immigration up to 11. No wonder they think it's a failure, the will of the people was ignored (and it's having real political blowback, with both major parties about to get reduced to dust according to polling).
I think you could argue Anthropic could be worth $1T. With AI becoming an essential work utility, every global knowledge worker would want a claude subscription. There are 650M to 1B of such office workers. 300M workers × $50/month × 12 = $180B/year. The genie is not going back into the bottle, I have seen what claude code can do when properly connected to tools.
But Micheals arguments are valid. There could be competition, or even local models, thus indeed becoming 'commoditized'.
What probability do you assign to that, especially since CC harness code leaked?
Because I used frontier models this weekend (I had 78% of my assigned tokens for this month left, I wanted to burn them before June 1st, ended up with 24% left), and tbh, I don't see much of the improvement compared to the models I use day-to-day. I'd rather pay less for a slightly worse model. Stacktrace analysis (or any bug analysis really) is where LLMs have the most success rate imho, and free models are good enough since last year. As for coding/architecture tasks, frontier models seems to hallucinate less, but I wonder if it's the guardrails or the he model themselves.
Well... $1 a day is not that far from that hypothetical $50 a month though.
Especially if it gives you access to significantly more powerful models (which it does).
EDIT: i still find absurd thinking that all those subscription would go to a single company, let me be clear. But that $50 price doesn't sound unreasonable at all.
> Especially if it gives you access to significantly more powerful models (which it does).
Anthropic and OpenAI are losing lots of money with their subscriptions. They are giving away access to those powerful models for cheap. The Deepseek price is the API price, which is the only sustainable approach here
AI is far from becoming an essential work utility, Anthropic will not be used by approximately 100% of the office workers in the developed world, and a price to sales ratio of over 5 for a company that is struggling to become profitable due to high operating expenses seems exceptionally high.
The problem with all these companies is that they are priced as if their training and inference costs are going to come way down, but somehow only for them specifically.
thing is, we have both local models and local hardware and a true evaluation would do a calculation before openai inflated thw market, before nvidia made circular deals and the other distortions.
i think youd find the ROI is nowhere near the API rates are the "price support" is entirely a figment of billionaires and their parasites trying to corner the market by horde logistics
reply