I learned that running goal for hours produces exponentially more slop than running targeted prompts over and over again manually.
Personally, I use gpt 5.5 high with planning every time and plan various smaller features/changes in parallel, then approve them one after another. This allows me to steer it (which I need more often than not) before approving the plan, thus reducing the otherwise accumulating slop.
Using goal doesn't work for everyone, unless you have an unreasonably strong test suite or harness that the agent can verify against.
It's funny, they sell you a subscription for frontier models, then over time begin to nerf them rapidly and no one talks about it. Should give me a discount when they reduce reasoning effort silently on the server side!
But on the other hand, I've been using 5.5-high on a daily basis in multithreading workflows, i.e. in parallel. I'm barely exhausting my weekly limits. I can't even Human-as-a-Service fast enough to catch up and read all the plans and implementations it does. So there is that.
I founded a UG with 2 friends. 7.500 capital, 2.500 each. From that money, we paid the notary. We drafted with chatgpt on our own and presented it to an attorney for review, ~300€. Notary ~1.200€. All in all, we are 1.5 years in, we still have ~3.000 left from that 7.500 capital. Obviously you're doing something wrong
> AI users were actually 19% slower, but they thought they were 20% faster.
I don't know who made those numbers up, but for me... I can almost certainly guarantuee, I have never been so relaxed before. Doing multiple paid projects simultaneously due to AI, still leaning back, customer's are happy. I can confidently say: if you know how to leverage it properly, you can be both more efficient and relaxed at the same time. I'd also argue, if you use a combination of SOTA models to code and review and put in some own thoughts, too, then code is also GG.
I was wondering "use" means here, as-in.. does it not recirculate? And apparently the answer seems to be: it's circulated/vaporized into the air. It may fall down as rain somewhere else, not necessarily in the local area where the water was withdrawn from, effectively draining the water from the local area at least.
Also, I was wondering, what does 2.5B gallons of water equate to? Here's the answer for curious minds:
> Using EPA’s cited 82 gallons per person per day figure, 2.5B gallons/year equals the annual household water use of about 83,500 people.
I did some further math... If 1bn users world wide leverage AWS services in their daily routine (netflix, whatever, ...), the formula becomes this one:
while unified memory may offer better performance than unsoldered DDR system memory, it still won't be as great as 1.8TB/s bandwidth on high end consumer GPUs right now.
nvidias master plan may be making it the new normal to have "only" 400GB/s bandwidth, thus gatekeeping local model usage further behind "more memory but not as fast as the cloud can do it"
For a little complex changes, I always run codex (5.5-high) in planning mode first.
I have linked various docs/{ARCHITECTURE,BACKEND-GUIDELINES,NESTJS-DI,..}.md etc. from AGENTS.md so they can quickly discover relevant docs at planning time, only if they are needed. No need to know react specific stuff when it's dealing with a backend problem for example. I typically blindly approve plans made by the agent with a fresh context, because that's as if I had prompted it. Works the best for me.
Using /goal however, it's really just constantly compacting and doing it's thing, of course it gets sloppy. If only there was a state machine that would transform tickets into a Planning Mode Prompt, then use, idk. guardian approvals (somehow a "Product Management Perspective Lens" approving or making changes to the plan) and then letting a less capable or less reasoning agent execute the plan, I think that would work the best.
> prompts with >272K input tokens are priced at 2x input and 1.5x output for the full session for standard, batch, and flex.
which is basically maxxed out quickly. So there is 2x (the first lever)
Then there is the /fast mode, which they state costs 2x more (for 1.5x speedup)
And then there is the model base price ($2.50 vs $1.75), well yeah thats 42% increase. It is in fact a 5.7x total increase of token cost in fast mode and large context. (Sorry for the confusion, I thought it was 8x because I thought gpt-5.3-codex was $1.25)
At this point, the pelican benchmark became so widely used that there must be high quality pelicans in the dataset, I presume. What about generating an okapi on a bicycle instead?
I set up windows 11 on a laptop for my dad so he can read emails and browse the web. Came back 3 months later when he told me he couldn't see the PDF files anymore. Turns out he installed THREE different PDF viewers that he randomly found on google, they installed tons of bloatware/spyware, replaced browser toolbars and searches etc. to a point where I decided to just restore from a recovery point. Told him not to download weird stuff (again) and ask me when he needs help.
At that point I questioned myself: I really should have installed linux for him.