sumitkumar
sumitkumar on Hacker News
friction is a necessary evil
It will hurt their cloud business as that model would be served by everyone and not just them.
This seems a lot like fishing. Cast a wide net with claude to find and solve an 'unsolved' problem. Given this, it is not enough to verify the problem and the solution but also the history of the problem and if it really existed or has simply been collectively hallucinated.
This is impressive as it is optimizing the effort on the low, but not too low hanging fruit.
"In late 2020s, while the whole world was focussed on AI, automation and resultant economy four major mathematical study branches were discovered by human researchers which took AI a long time to catch up with"
The libraries are there for avoiding repetition. It is useful on its own without other benefits like implicit/explicit boundaries, readability by humans etc. If all lego(base unit software) pieces were unique shaped it might fit your requirements more efficiently but the cost of building those unique shaped legos explodes with the size of the lego piece required.
LLMs compress the known shapes well and fit them to solve for a problem but they are still not good enough to build from scratch a large new lego piece which fits a full problem perfectly. And such a large lego piece might not be the most efficient solution either and might be difficult to prove so.
It is based on Anthropic Uptimes. If anthropic wanted 99.999 they will have to acquire 10 times more GPUs/infra to match the traditional free resources of 99.999 services.
Agree. Just pay attention to the follow up questions a learner is asking to see the progress. If the follow up is just "continue", "go on", "next" or a non-sequitur then it is smell of a stall. If it is challenging or filling a gap in the answer then it is progress. So production from the learner is the only signal of worth here not the quality of LLM response, the time spent or the ability of the learner to reproduce the facts given by the LLM.
With human devs, errors become recovery instructions.
400 Bad Request is fine for a client developer who reads it once at design time and fixes the code forever. it's dead weight for an agent that must self-correct from the string alone. "Expected ISO-8601, got 03/04/2025" is now a functional part of the interface.
Not all agents are developers who can code their own interface.
Google improved chrome by reading code: Google built an agent harness using Gemini to scan Chrome's codebase, trained on a knowledge base of prior CVEs and the entire Git history, with a "critic" agent consuming developer-supplied SECURITY.md files.
In my experience, I find it to be exceptionally good at exploring and fixing the edge cases. Of course the output is not human maintainable for these fixes and needs to be heavily tests controlled, refactored or just accepted as being agent-maintained going forward.
electricity outage and battery running out is the end game for any real prolonged external emergency. Internet connection is just the soft edge.
And the broken glass breaks the dishwasher. It is a feature: planned obsolescence.
Can you please give some examples of the best tools/businesses which are being built like this? Genuinely curious. I know the vibe tech is really good but I don't know of any upstarts using it, everyone cool is building/distributing the vibe tech. There are some games/copycats but nothing production grade in my radar.
Well, thank you for clarifying. The signal is getting lost in the noise. I assumed too soon after looking around just the datasette org github account and seeing so many repos and code being built so fast.
Our leader is Boris Cherny.
Simon needs to resist the pelicans(and the django mindset) and Garry needs a new loop which can loop on itself without any human trigger so that the agents can "dream" better. Who knew that it was not just the models which could hallucinate.
I just went through the github project repository.
It has 119 repositories.
Is this how AI slop looks like in code? Made for the agents, by the agents?
Is this separation of concerns or context management with agents as a first class residents and humans merely acting as custodians?
Not that far. Lawyers had a great deal of influence in creation of all modern nation-states, human rights, international law and maintenance of the core social contract in the modern society.
Similarly lawyers/bankers were the ones who built in trust in capital, contracts, businesses and protection of investor rights. Delaware c corp is not an outcome of bad guys.
This happens in any industry where value/status are at a premium.
Finance, Law, VC guys were good too in the beginning but when the value/status change happens it attracts certain kind of guys who are average in talent but excel in demonstrating value and social management of the value/status.
Another change which has happened recently is that the economics of engagement farming have become common place wisdom as already proven effective for everything from selling books, personal brand, career skill/virtue signalling, staying relevant.
Due to this everyone is talking more without restraint and not keeping in their own lane of earned expertise.
Everything is extractive. Farmer plants seeds, partially sets the environment. The work is done by the seed/sun/soil/water. And so is every profession: labour or not. Most of the business are structured in such a way that someone can exploit them to make even more money. The whole vendors and b2b system is mutual extraction.
Looking through wages and trying to find a ceiling(by time/effort) on the value creation by a human is one dimensional at best.
Documentation is worth it only if it is read. If your coworkers don't read/remember/respect the documentation process then people tend to not keep the docs up to date. Unless the docs are for users who you don't want to come to you at all for support.
Claude is a better reader. I have to just tell it to read the docs/specs sometimes.
True, so the interference is the "computation"(heavy emphasis on quotes) which gives rise to the principle.
By Fermat's principle, a ray of light has to know where it will ultimately end up before it can choose the direction to begin moving in.
So either something is computing it or some exploration is happening at quantum level and we just see the final result.
Yes this is more like compression to remember and not for learning/understanding.
The original did not come out of a vacuum. It was done on multiple generations of meat. Even though this one uses a little bit of silicon, it is still standing on the same shoulders.
The weights start with a random manifold.
The training takes data and shapes the manifold, weight by weight, in many cycles.
Once the training is the done manifold is fixed.
When a new inference has to be done the query(q) is projected in the manifold space.
This projection is dropped on the manifold and the gravity of the manifold gives an answer of q+1 length.
Which(qw+i) is dropped qw+n times to output a final response of n length.
The gravity is created by repeated multiplication(of the weights/input) to find out how the projected embeddings should fall according to the manifold in the GPU.
so can we all agree that LLM models/agents are bad at BFS for exploring a problem space but are good at DFS to implement a solution if the context/requirements are rich enough.
The problem is because of the RL and system prompts by the providers which tend to placate the user using certain language tones and register for response. This objectively messes up the generation while steering it into acceptable responses.
Most of the conversational skill and perceived intelligence of these models in hidden in RL/system prompts.
But the general use case is not the most efficient for a greenfield to-be fully managed by an agentic system code-base. It is built to be good around the scaffold(programming like humans) and not the actual problem space.
Anthropic's target should be a codebase designed for agentic comprehension from the first commit. Here the codebase adapts to the agent. You can enforce conventions, structured metadata, semantic indexing, explicit dependency graphs. Whatever makes the agent's job trivial rather than heroic.
He would be right if claude code was written by a team of humans. The AI written blob is slowing progress.
But once a human learns a function their errors are more predictable. And they can predict their own error before an operation and escalate or seek outside review/advice.
For e.g. ask any model "which class of problems and domains do you have a high error rate in?".