I've always thought that extensive throat-clearing and prefixing the Treaties of Westphalia-length instructions into the context window was unnecessarily baroque when you can just talk to the agent.
I guess part of it is also that I don't mind doing 'hand-edits' like for example LLMs love to say "// so and so removed" I just go and remove that manually later rather than being like "don't comment about what you removed!11" cause you're really fighting deep grooves in the model's behavior at that point.
But I also have a hands-on human-in-the-loop working style so I guess maybe for people who just want to say "implement all open features in github issues" and walk away maybe there needs to be more of all this CLAUDE.md stuff
However I suspect there was always some gearhead type attraction to setting up detailed harness configs that may be unnecessary and more like hobbyist tinkering.
Very interesting. The past few days' trend is so common in small talk and general weather discourse ("it was warmer last week") but underrated in weather apps. It's helpful that Open-Meteo provides that data
This output is incidentally perfectly context engineered for LLMs too; I pasted the Delhi page contents into Arena .ai and asked "what's the gist" and qwen3.6-plus said (among other things) " A hot, humid day with a slight chance of afternoon showers." and gemini-3.1-pro-preview said "Mostly clear, but hot and very humid. While the actual high is 34°C, it will feel like 42°C in the afternoon."
As of early June 2026, Opus 4.8 in fast mode cost $50/M output tokens and Opus 4.6 & 4.7 cost $150/M output tokens in fast mode
How can supply and demand explain the price drop? Was it cheaper to serve Opus 4.8? Is the demand for the newer Opus lower than for the older Opus? These are just fixed prices that seem picked out of thin air
One of the strangest things in the AI industry is 'tokenomics'. It's not very clear why using GPT-4 in early 2023 was so expensive and then six months later 20 bucks could get you a fair amount of GPT-4 inference. This pattern has continued across various labs/providers for years--there is a continuous see-saw of pricing that doesn't seem related to anything.
So what open weight models do is at least provide a baseline of inference cost to add some sanity to the price markers. And of course predictability too--if you really want Kimi K2 instead of K3 you can still use it.
So the competitive pressure and predictability offered by open models is helpful for users
I think it depends on the context too. Dale Carnegie (in his famous Win Friends.. book) mentioned how it's annoying to receive letters that say Dictated But Not Read. So people don't want to read (for example) a twitter reply or email that's written on their behalf rather than their own thoughts
When it comes to something like a LinkedIn post however who cares. Everyone uses AI cause the content itself is mercenary anyway right. Thought leadership.
But I have been using Sonnet to write some Youtube narrations and I think I've discovered a kind of value that is not 'AI slop'. I already posted one of them -- about the giant pink hologram in Blade Runner 2049 -- and it's been getting thousands of views and a dozens of likes when clipped into shorts. [https://www.youtube.com/watch?v=sK-TsLQkIos]
The text is like: "The neon is pink. Her hair is blue. Her face is light. She smiles because she was made to smile, and somewhere beneath the awareness of that fact is the older, simpler, more animal truth: a face is smiling at you. Someone said your name. Someone said “you look lonely” and meant it as an invitation rather than a wound."
To me that is a kind of LLM writing that's "value above replacement". Like if I lost that Sonnet 4.6 script I wouldn't be able to get it again it was just a particular permutation of the neural network manifold that emerged from our discussion about the movie and the style I wanted etc.
This is like the late 80s/early 90s cyberspace idea... Imagine you're walking through a 3D ramp and all these buttons are yelling at you from all over the atmosphere... Login here! Start here! Contact us...
I think this nostalgia is useful for reminding us there are aesthetic forms beyond what Jony Ive et al happened to design and became the dominant form factor. The 2000s explosion of MP3 players had a lot of little ideas like the iRiver Clix had 'press the side of the device' as a navigation system
Works in Claude.ai, Claude iOS, Claude Code, OpenAI Codex, and any MCP-compatible client.
It adds a handful of tools eg
clock_get
clock_day_info
clock_convert
clock_convert_unixtime
clock_delta_utc
This way you can ask "what time will it be in Seattle when it's 5pm in Tokyo today" or "I'm halfway through my slides how much time is left to my 5pm deadline" etc
The idea in all the examples in this page seems to be: write a string with some code and then write some more code that processes the string in a way that when the string is finally output it contains the variable assignment (of the string), the variable contents, and the later program code.
Indeed that is what the first wikipedia example does
a = 'a = {}{}{}; print(a.format(chr(39), a, chr(39)))'; print(a.format(chr(39), a, chr(39)))
I dunno, my explanation seems to make more sense in my head than the analogy this article fixates on of a machine that can build a machine etc
Very interesting that one of the components is "AA-Omniscience Index"
AA-Omniscience Index (higher is better) measures knowledge reliability and hallucination. It rewards correct answers, penalizes hallucinations, and has no penalty for refusing to answer.
This seems to be a good proxy for param size/density and the ranking breaks down as such: Claude Fable 5 (with fallback), Gemini 3.1 Pro Preview, Claude Opus 5 (Max), Grok 4.6 (high), Gemini 3.6 Flash, GPT 5.6 Sol (Max)
I've thought for a while that Gemini 3.x has 'big model smell'
Very interesting. Here in India people were very concerned about potential cooking gas shortages (LPG) when the disruptions began which is also a good example of usually-overlooked dependencies on the the Strait
I think one way to look at this is that--why are Meta and Google so successful? Because they are basically an index fund on the world economy right. If you want to sell soup in Jakarta you advertise on Instagram. If you want to sell paint in France you advertise on Youtube. If you want to sell T-Shirts in Boston you advertise on Facebook
So then the flipside is that without ads these sellers now have to complete head on with giant companies who have all the brand recognition. Now you need to somehow make your Etsy store viral instead of just putting it in Facebook ads
I think there is a fundamental issue here of what building software even means
If you think you can just assign Github tickets to AI agents and go drink daiquiris on the beach I think you'll find that you end up with more and more towers of abstraction and indirection. There are 'points of view' that emerge during coding I think. And at some point you as a human have to be like "wait... what if we use Redis here". "Wait.. the API is already returning the data we need". "Wait... let's not add customers to the report who have not been active in the past year". Stuff like that
Have you ever tried analyzing a CSV in an LLM? Maybe then you'd get what I'm saying
Heck for a lot of questions the LLM won't even write Pandas and just try to eyeball the arithmetic and be like "Yeah March 3 to March 7 values add up to..."
An absurd example is how harnesses like Codex, Claude code have no 'cut/paste this range of chars' affordance so if you want to move a block of code the LLM galaxy brain rewrites it in token space in the new file.
It ~can~ be done using standard bash utils but there just isn't a kind of standard pattern for 'move this text block via CLI tools' so the LLM and human both overlook it
Yeah I guess the dichotomy between 'tools' and 'code mode' isn't really as clear especially if the tools themselves are intelligent and truncate long results and have start_range, end_range type arguments. So the gist is about 'what part of the output needs to reach the LLM' and a 'code mode' approach can help orchestrate that just like good tools can
Yeah, of course. The MCP server tools do use Python in the backend but just look at the affordances it provided in the args -- named columns, a fixed match/replace API, a way to specify the range of rows. The LLM is not gonna create a little library with these affordances in an ephemeral REPL it's just gonna YOLO some pandas code that you hope did what you wanted and addressed the right columns, had the right types, etc
http://twitter.com/firasd
firasd at gmail