As by this gameplay video, after the water palace walking at this place should spawn some goblins that captures the hero, but in my game they are not spawning.
I'm playing from the Nintendo Switch Collection of Mana.
Any idea what am I missing?
I also explored further, there's that missing girl, with a boss of 2 enemies that heal themselves. I suppose I need one more in my party before being able do defeat them.
Looks like the powders have no use, as the crafting starts with the Ore.
On my second playthrough (9 years ago) I avoided looting everything from everywhere. And it actually made the game better for me, because things like Gourmet skill now had a real use, and a pub selling food was actually something interesting.
Now on my current playthrough I had forgot that… looted everything again… and food has no much meaning anymore. Also the Blood and Wine woman cooking for me will have no meaning. Kind of sad.
Fable is fantastic, best ever!
But I just noticed after around 14 requests on the same conversion the quality degrades a lot. I don’t know in terms of session context how much it is, but after some requests looks like I was talking to gpt 3.
Anyone else?
How crazy it is that Sonnet 4 for on the Coding Avarage column of livebench is better than Sonnet 4.6 and better than Sonnet 5?
https://livebench.ai/sort=Coding+Average&highunseenbias=true
GPT 5.5 xhigh answer (that I probably should have used medium):
Many problems look complex because they have many words, but the answer is basically a known pattern. Humanity has heroically invented verbosity to disguise “use the obvious thing.”
The rule of thumb:
Use maximum reasoning when the hard part is reasoning. Use average reasoning when the hard part is recognizing a standard pattern, checking facts, or applying common sense.
Here are examples where “maximum reasoning” feels tempting but probably gives the same answer.
Problem that feels complex
Why average reasoning is enough
Likely answer
“I need to choose between PostgreSQL, MongoDB, DynamoDB, Firebase, and SQLite for a normal app with users, payments, settings, and reports.”
It sounds architectural, but it is a standard CRUD/product app question.
Use PostgreSQL unless you have a very specific reason not to.
“My JavaScript code has async bugs across several functions and sometimes returns stale data.”
Most of these are known patterns: missing await, race condition, stale closure, bad dependency array, no cancellation.
Use better async control, cancel stale requests, add logging/tests.
“Should I split my backend into microservices?”
If you are asking this early, the answer is usually boring and correct.
No. Start with a modular monolith. Split later if pain becomes measurable.
“I have a complicated personal decision with many pros and cons.”
Often the AI cannot infer your real priorities better by thinking harder.
Clarify values, constraints, risks, and reversible vs irreversible choices.
The big traps:
1. Lots of constraints do not always mean deep reasoning is needed.
Sometimes 18 constraints collapse into one obvious blocker, because apparently humans enjoy making shopping lists for decisions already settled by constraint number three.
2. Missing information is not solved by stronger reasoning.
If the answer depends on current laws, prices, product specs, medical guidance, APIs, documentation, schedules, or local rules, the right move is fresh information, not maximum reasoning.
3. Standard engineering questions usually have standard answers.
Postgres, boring monoliths, managed auth, profiling before optimizing, tests before refactors. The classics are classics because they keep civilization barely operational.
4. Subjective decisions need preference clarity, not more compute.
For “best game,” “best city,” “best laptop,” “best stack,” the AI needs your priorities. Max reasoning cannot discover your soul, which is inconvenient but probably good for everyone.
Where maximum reasoning actually helps:
Multi-step logic with hidden dependencies.
Debugging across several files/systems.
Tradeoff analysis where each option affects the others.
Architecture decisions with unusual scale, compliance, latency, cost, or team constraints.
Math/proofs/algorithms where a wrong intermediate step ruins the answer.
Legal/medical/financial analysis where careful caveats and source-checking matter, though experts still matter because reality remains annoyingly non-automatable.
A decent heuristic:
Use average reasoning first. If the answer feels generic, misses constraints, or fails on edge cases, then escalate to maximum reasoning with the exact failure mode.
I asked why its responses are so less complete then other top models, and it answered me it has system prompts telling it to be short and concise. I feel cheated, I had a lot of trouble here until I found out this. Now I’m trying to workaround these unsolicited instructions.
Did you knew about this?
I’ve been using GPT Codex heavily and things start falling apart around 180k tokens. Sloppier reasoning, dropped context, wrong assumptions. I capped my sessions at ~170k, which helps, but now compaction kicks in constantly. Where quality actually gets flaky in real coding work for you? And how do you deal with it?
I’m getting my usual software development with gpt 5.4 with some lazy results this week. Anyone else with same impression?
It was great until this week.
The infographic version shows the full structure: Saturn’s worldline as a thick golden tube, the rings as a continuous helical sheet of light made of billions of fine particle threads spiraling alongside, and the major moons (Titan, Enceladus, Mimas, Rhea, Iapetus) as thinner helixes coiling around at different radii and speeds. One full orbital cycle = ~29.46 Earth years.
The close-up version you can see Titan as the thick golden braid on the left, smaller moons as fine threads, and the ring structure as those translucent flat sheets wrapping the central tube.
Saturn’s rings are only ~100-400 million years old and will probably dissipate in another ~100 million. So in really deep time, the ring ribbon would only exist as a bright temporary section in the middle of Saturn’s worldline, a 200-million-year flash of beauty in a ~4.6 billion year worldline.
I watched this video about the Möbius strip and Klein bottle, which uses time to help visualize the 4th dimension:
It got stuck in my head. If every moment is just a 3D slice of a larger 4D object, then the solar system as a single spacetime object should look very interesting.
So I went down a rabbit hole generating visualizations to explore the idea. The Sun becomes a long tube extending through time. Each planet becomes a helix coiling around that tube. And, in 4D, the Earth and Moon become a single braided structure, like cosmic DNA. The Moon oscillates around Earth's orbit about 12 times per year, and once you materialize that motion through time, you literally get a double helix.
Made me think about how much of what we call "objects" are really just slices of much larger 4d structures we can't perceive directly, because our consciousness only sees one 3d moment at a time.
Generated with AI after a lot of iteration on the prompts, most models default to "flat spiral galaxy" and you have to fight them quite a bit to get the 4D interpretation.
I’m wondering if Anthropic is injecting all memory each request and/or doing the “generate memory” in a way that spends my tokens faster.
Has anyone tried delete all memories and disable this to see if is there any difference?
Hello!
I’m a software developer thinking about how to keep all my LLM conversations in one app instead of having them scattered across ChatGPT, Claude, Gemini, etc.
Ideally, I’d like something where conversations are stored locally, preferably as Markdown files, organized in folders/projects, searchable, and not locked into a single provider or model.
And of course, using my subscriptions (talk to claude code and openai-cli/codex when possible, not only using api keys).
Later I might want to send the same message to multiple models and compare the answers, but that’s not the main goal right now. For now, I mainly want something like “Obsidian for LLM chats.”
Maybe fork and adapt LibreChat, but I’m wondering if there is already something closer to this idea.
Has anyone tried or started something like this?