Pipelines are often dynamic, how is this achieved?
Pipelines are just a description of computation, sometimes it makes sense to increase throughput, instead of low latency, by batching, is execution separate from the pipeline definition?
It's quite tricky as they optimize the agent loop, similar to codex.
It's probably not enough to have answer-prompt -> tool call -> result critic -> apply or refine, there might be a specific thing they're doing when they fine tune the loop to the model, or they might even train the model to improve the existing loop.
You would have to first look at their agent loop and then code it up from scratch.
I bet you could derive a lot by using a packet sniffer while using CC and just watch the calls go back and forth to the LLM API. In every api request you'll get the full prompt (system prompt aside) and they can't offload all the magic to the server side because tool calls have to be done locally. Also, LLMs can probably de-minimize the minimized Javascript in the CC client so you can inspect the source too.
edit: There's a tool, i haven't used it in forever, i think it was netsaint(?) that let you sniff https in clear text with some kind of proxy. The enabling requirement is sniffing traffic on localhost iirc which would be the case with CC
I would say IMO results demonstrated that. Silver was tiny 3B model.
All of our theorem provers had no way to approach silver medal performance despite decades of algorithmic leaps.
Learning stage for transformers has a while ago demonstrated some insanely good distributed jumps into good areas of combinatorial structures. Inference is just much faster than inference of algorithms that aren’t heavily informed by data.
It’s just a fully different distributed algorithm where we can’t probably even extract one working piece without breaking the performance of the whole.
World/word model is just not the case there. Gradient descent obviously landed to a distributed representation of an algorithm that does search.
Most software is just middlemen collecting money. Thats the reason why there is no economic drastic growth even with 2-3 years of AI.
What you’re worrying about is just bigger middlemen.
If you were not a value extracting middleman there would be no fear of replacement, because you can always create more than what you take.
I’m glad if this causes a shift in the industry and we lose x analysts, x architects, x scientists, data engineers and all other formulaic middlemen that just live in a weird middlemen economy.
It is immense luck that we don’t have to actually produce something and we get paid but it is much better if we’re forced to actually do something that isn’t empty.
If by features you mean tracking state per user, that stuff can be tracked without Flink insanely fast with Redis as well.
If you re saying they dont have to load data to update the state, I dont see how massive these states are to require inmemory updates, and if so, you could just do inmemory updates without Flink.
Similarly, any consumer will have to deal with batches of users and pipelining.
Flink is just a bottleneck.
If they actually use Flink for this, its not the moat.
Yea, the Monolith paper by Bytedance uses Flink but they only say it's in use for their B2B ecommerce optimization system. Maybe this is intentional ambiguity, but I'd believe that they wouldn't rely on something like Flink for their core TikTok infrastructure.
My hunch is we start to learn a lot more about the core internals as Oracle tries to market to B2B customers, as Oracle is wont to do!
Flink is not really a performance choice, it's bloat to throw software as fast as possible at problems. I don't think there's any benchmark demonstrating insane capabilities per machine. I definitely couldn't get it to any numbers I liked, given other stream processing / state processing engines that exist (if compute and inmemory state management is the goal). Pretty sure any pathway that touches RocksDB slows everything down to 1-10k events per second, if not less.
The problem of finding out which video is next, by immediately taking into account the recent user context (and other user context) is completely unrelated to what Flink does -- exactly-once state consistency, distributed checkpoints, recovery, event-time semantics, large keyed state. I would even say you don't want a solution to any of the problems Flink solves, you want to avoid having these problems.
I tried just repeating guó for as many times as symbols and repetition was not recognized.
Although I like the active aspect of the approach. Language apps where sound is the main form of learning should have a great advantage, as any written text just confuses as every country has its own spin on orthography. Even pinyin, despite making sense, for a beginner, has so many conflicting symbols.
I think this is standard. It applies to domains as well. I experienced government services blocks as well -- they send me an email, yet block my reply. I complain every time and rarely does anyone care, the support person does not escalate, so my email remains blocked, sometimes I'm told system is working as configured, completely ignoring that I am a real person and system is hostile towards me.
It's just general fragility of tech and lack of care from the creators/maintainers. These systems are steampunk, fragile contraptions that no one cares to actually make human friendly or are built on crappy foundations.
This has nothing to do with decentralized networks. It's simple incompetence.
If you haven't received any mail from a mail system before (or in a long time) and then it sends you one message, it probably isn't spam, because spammers are typically going to send you a large number of messages. You also typically want to let the first few messages through so the recipient can see them and then classify it as spam or not, so that you get some data on how to treat future messages from that sender.
This is the same thing a centralized system should be doing with individual users. You impose some reputation on accounts (e.g. by sender/registration IP address) and then if that address starts spamming people it gets blocked, and otherwise it doesn't.
This is one of the things that E-Delivery (something which Europe is now implementing[1,2,3]) is going to fix.
It's sort of like email, but based on the XML stack (SOAP / WSDL / XML Crypto / XML Sig), with proper citizen authentication and cryptographically-signed proof of sending and delivery.
We are repeating obvious things here aren't we? I moved to Germany from a very pro IT country Finland. I've been here now for 15 years, and while I still disagree with their idea of dismissing email, I kind of got used to it. A couple more decades and it'll happen...
The main issue is that who is supposed to implement it? The gov has 2 possibilities: hire a contractor, or do it themself. DIY has the issue that nobody wants to work for the gov because as any IT specialist you'd earn 1/3 or 1/4 of what you would earn in a private company. Stateworkers here cannot be fired. So you trade money for extreme "stability" (read: laziness). Hiring a contractor requires money they also don't see the necessity to spend. And that's how you end up in this situation. There are also other issues like no national wide implementation plan. Every state, every commune has to figure out and build stuff themself.
This internal compulsion is just learned behavior. The society conditions you to work instead of play.
Nothing wrong with that, I have that compulsion as well.
Having a compulsion to play, purely for the sake of playing is a much healthier view. Useful, not useful, hard problem, easy problem, should not matter, you're playing.
Sometimes you can't be useful, yet you can always play.
All stems from inability to have systems without labor. Work, work.
I like how Pope John Paul II flipped the narrative and said work exist for the person, as a way for person to express itself. Made me realize how even communism stays trapped in labor mentality.
As we mature mentally, we need more interesting games to play, more interesting challenges. Work is many times the result of this.
It's the same with romance. When we are children we have a crush on somebody, become pretend "boyfriend and girlfriend", and as we mature the game becomes more interesting as it becomes real.
But it's all a game throughout life.
So perhaps it is those who enjoy work who has elevated their spiritual level, and not the other way around?
Was about to comment the same till I found your comment.
I have this compulsion too, and did some deep-diving at some point through therapy. I found that really it's just likely conditioning from family/society.
If you are generally praised for helping out whilst growing up and this is when you receive a lot of love/attention, it's natural to build pathways that favour this and thus behavioural patterns.
I like this thought. It is interesting to look at our current societal/economic systems on the earth and realize none of them will survive the death of scarcity.
In abstract terms capitalism doesn't depend on scarcity? Capitalism as in centralisation of the means of production (even when there is no human labor anymore).
It’s funny how there is continuous reinvention of parsing approaches.
Why isn’t there already some parser generator with vector instructions, pgo, low stack usage. Just endless rewrites of recursive descent with caching optimizations sprinkled when needed.
Hardware also changes across time, so while something that was initially fast, people with new hardware tries it, finds it now so fast for them, then create their own "fast X". Fast forward 10 more years, someone with new hardware finds that, "huh why isn't it using extension Y" and now we have three libraries all called "Fast X".
Because you have to learn how to use any given parser generator, naive code is easy to write, and there are tons of applications for parsing that aren't really performance critical.
Steps 1 and 3 are heavily dependent on the data types that make the most sense for the previous (lexing) and next (semantic analysis) phases of the compiler. There is no one Token type that works for every language, nor one AST type.
The recognizing the grammar part is relatively easy, but since so much of the code is consuming and producing datatypes that are unique to a given implementation, it's hard to have very high performance reusable libraries.
Pipelines are just a description of computation, sometimes it makes sense to increase throughput, instead of low latency, by batching, is execution separate from the pipeline definition?
reply