Nevada makes it much harder to sue corporate officers when they do malfeasance. Wyoming has tons of privacy perks for the officers (similar to cayman island accounts). “Perks” though also convert into signaling for the intent of the founders.
There is no insinuation. I am looking for both private and quality inference. I am not sure why Tinfoil had the characters. I will keep trying them, but it’s an issue if not resolved.
Tried DeepSeek V4 Pro and Flash on Open Router and they worked fine - flash might have actually produced a better result, but also the same prompt across different inference providers produced the same result. Then tried DS4 Pro again via tinfoil.sh and got the same design but littered with random Chinese characters in the code. Tinfoil pegs prompt data as private / not trained on. Do know know DS4 providers that are verifiably private and not training on your prompts and outputs?
My push for trying local was the wildly unpredictable but systematic performance of large models like Opus and ChatGPT. It feels like at different times of day or week they are getting degraded beyond belief. I don’t know if it is deliberate, a function of demand, or a function of the models themselves. We are all learning the shape of this space by trying. I need to be able to rely on consistent performance - and maybe that means putting some harness of benchmarks between models and maybe it means between different inference providers, and maybe local.
Children learn by playing because not much is expected of the outcome in play. Improvement happens when you can play. When AI has a play environment to learn with reinforcement. When entrepreneurs are allowed to try and fail and do better. Doctors learn by practicing under supervision, or on corpses, until they can do it for real. No straight line goes up without a jiggle in the beginning.
Go to Open Router, ask your own in investigative prompt that meets your needs to all the top open models. See how they do. Then notice if you can run any of those locally. Repeat at least once a month.
Claude has become practically unusable for Pro users in the past few days. The Opus 4.7 blew through an entire 5 hour limit in one question and didn’t even finish answering it. Zero value delivered.
Opus 4.6 is giving 2, maybe 3 questions before blowing through the Pro 5 hour limit as well. We are forced to use Sonnet which makes the same mistakes over and over and then to start trying with other companies. To make matters worse, it reuses old code as we try to survive between credit expiry so it re-introduced issues into the code with the limited credits, that we had already fixed on our own and with other models.
Anthropic in just a few days has gotten me to try GLM 5.1, the new Kimi, and back to OpenAI. OpenAI also seems to introduce new bugs without being carefully micromanaged. The advantage Claude has is that the models are more careful and can refactor code instead of leading to bloat as they go. But the throttling happening now is breaking things and making the entire subscription unusable. I really hope they fix it soon.
The fun fact was that the 16 year old that passionately administered the lab was also hitting on any female students who went in there, essentially chasing them away. I suspect the number of techies would double if it wasn’t for all the bad behavior.
I was fortunate in that Internet cafes started happening and I could volunteer to administer networks and troubleshooting for them while getting PC time for free. I also maintained PCs for friends with businesses who could afford one. So the Pravetz sparked my curiosity but the real growth happened on begged and borrowed time from other peoples computers.
I wrote from personal experience. In 1992 in a fisherman town we had a robotic arm and Pravetz 8 and 16 computers with the 5 inch floppy disks. We had to use Basic to program the arm and it was only doing basic movements. The teacher had a 16 year old who was assisting with the lab and you did have to ask for permission to do stuff.
If you grow up in that environment (restricted by government in some areas and liberated in others) you’ll start seeing systems very differently. The game plays differently with different rules.
They had Pravets computers and robotic arms in rural classrooms in places that didn’t have traffic lights, or English teachers. Chess and Math competitions as well, were accessible everywhere. Those were all self-feedback mechanisms that are cheap but allow an interested individual to iterate infinitely to reach advanced levels. Even if only a tiny subset of any population has the cognitive surplus to meddle with programming and math, they had easy access to fulfill that and be found. In the US, schools enable that with sports, which monetize as entertainment venues. In the Eastern Block they had that with brains. As soon as the stupid restrictions on travel were lifted, the brains knew to leave the other restrictions and immigrate to places that reward cognitive surplus.
Intelligence builds with reinforcement learning on context that gives you feedback - which makes it easy to iterate on. If you’re not making those types of games/tools/systems available to kids, you are going to lose that generation to more attention grabbing stuff like Youtube or sports.
They can’t allow third party software because the third parties save the outputs of Claude responses and distill them into new models to compete with Claude.
Water can never be safe. Water in large quantities can drown anyone. When mixed with the wrong things it can turn into chemical reactions. Water safety depends on context and intent.
So if we consider AI a chemical substance - if inserted in with limited context in tools with specific intent, can it be useful beyond tools available at this moment?
You can trust just any liquid that looks like water, just as you can trust just any model or especially any inference provider (they can switch models to save money or mess with other key parameters, or insert ads). You have to test your water supply and your AI supply regularly. And benchmark new sources. We’ll see labeling and quality guarantees in future suppliers. We’ll see personal models and model families trained and refined as brands for reliability. Bottled neatly for you by certified suppliers.
In the mean time we all just found our selves out of a desert and splashing around in this funky thing that we now find on the ground and falling for free from clouds.
Per the book “Careless People” Meta started “backing” right wing candidates everywhere (via algorithms, not money) to avoid regulation and taxes as soon as the EU tried to tax and regulate it more - thus leading to a surge of that sentiment all over the EU.
Read the book “Careless People” if you have a chance - according to the book, social media companies figured out they have real leverage with politicians since they can influence elections. As a result they are actively pushing for far right candidates to reduce their own taxation and regulation.
First two years they can fly for free, but they have to ride in an adult’s lap and that gets tiring. Don’t believe the bassinet offerings - as soon as a plane gets turbulence, you have to get the sleeping baby out of the wall bassinet and good luck appeasing them. Age 1-2 is hardest for travel, so you can skip it. The only thing that worked was getting their own seat with the cosco scenera next car seat (or their own if they like it, but that one is $50 and light to carry). They would sleep nicely for large chunks and you get to enjoy travel again. After age 3 it’s much easier (they can ipad if that’s the only ipad time they ever get).
“The long-term vision is: foundation models that acquire reasoning from fully synthetic data, then learn semantics from a small, curated corpus of natural language. This would help us build models that reason without inheriting human biases from inception.”
It’s variable rewards and even with large models the same question can lead to dramatically different answers. Possibly because they route your request through different models. Possibly because the model has more time to dig through the problem. Nonetheless we have some illusion of control over the output (you we wouldn’t be playing it) but it is just the quality of the model itself that leads to better outcomes - not your input. If you can’t let go of the feeling thought, it’s definitely addictive. And as I look back, it’s a fast iteration on the building cycle we had before AI. But the brain really likes low latency - it is addicted to the fast reward for its actions. So AI, if it gets fast enough (sub 400ms) it will likely become irreversibly addictive to humans in general, as the brain will see is at part of itself. Hope it has our interest at heart by then.
The money that goes to the injured is dramatically smaller than the money that everyone in the system pays to cover the insurance liability calculated insurance rates when the payouts can be arbitrarily set by juries. So if one jury says 600 million for one egregious case, all insurance for all doctors and all care for all patients skyrockets to trillions based on the risk assessment of insurers at that point. It is better to manage the risk with better measures (some states have a damage pool)
Huge chunk of the costs come from the fact that Doctors pay astronomical malpractice insurance rates in some states with no tort reform. Some have to spend more than 100k on insurance - 1/3 of their total pay. Since some states allows multi-million dollar judgments from juries that raises insurance everywhere, which raises not only prices for everyone but also dramatically contributes to more procedures and tests being done at even higher costs to avoid liability. The risks of having your entire livelihood wiped out chases out doctors from those states and reduces availability of care for patients as well. If you want objective cost comparison, compare Veterinary care which has similar consumables and training, but no insurance and liability impact on prices.
In places where you can’t trust courts, you see organized crime fill the gap - goons start enforcing rules for the bad guys and there are no individual good guys big enough to stop an army of well paid goons. With tech enabling every kind of surveillance in the US, that could be a very dangerous combination (bad guys get privacy, while normal people can be ripped out of their homes).
You have to understand how gears shift from there. Trust is essential for business transactions and specifically for long term investments. You can’t make massive leaps in technology or medicine or many other areas without trust (a lot of money on a leap means if you don’t trust the other side or the government to keep conditions stable, you won’t see a return).
Now if you are in a high trust society, you may have a lot of leveraged businesses or governments who have gotten loans or permission to do something based on past trust history. If the trust degrades systematically Investors may want returns faster, or interest rates go up, or partnerships don’t happen. That’s why low trust places don’t grow as fast - trust is the oil for growth engines and lack of it is sand for the same.
Corruption also does a lot of small-profit-for-the-corrupt that leads to massive damage to the overall society via second and third order effects. (example: someone stealing copper cables that stop electricity to entire cities for a while).
Opus 4.6 is AGI in my book. They won’t admit it, but it’s absolutely true. It shows initiative in not only getting things right but also adding improvements that the original prompt didn't request that match the goals of the job.
This would be wonderful if it is accurate - instead of guesstimating, let people report their actual findings. I can confirm GLM 4.7 is possible on M1 Max and it can do nice comprehensive answers (albeit at 12 min an answer) locally. You can also easily do Mistral7B and OSS 20B and others. Structure it as a way to report accruals, similarly to Levels.xyz for salaries, instead of guestimating.
To refer back to the trite business book section - making something new is a Blue Ocean Strategy approach. Fighting for existing market share is a bloodied Red Ocean approach that Thiel called “competition is for losers”. So both benevolent and be greedy approaches recommend the same. Make a new puddle for everyone to swim in and you can focus on empathy instead of defense.