My AI researcher bot test is probably the most boring one I've ever run. I built a bot that reads research papers, mixes ideas, checks patents for clues, and so on, but it overthinks like crazy. I mean, it's great to see it being super critical with pretty much a 0% hallucination painfully slow. I asked it to find something it could solve quickly that would make people on X go "wow"... Guess how long it's taken so far? Almost 3 hours. In those 3 hours, it generated 31 hypotheses that might fit my criteria, dismissed a couple as "questionable", and concluded the rest were impossible to solve. Sheesh. I'm wondering if I'll need to upgrade to the $200 plan to keep testing this, though that feels a bit too steep for me. Then again, I could always task it with figuring out how to pay for itself 😅

I don't understand how this happened, but this happened. 40% of the 5 hour session limit, but 0% of the weekly limit, strange.

Really strange & it keeps going this way

After 40+ days of grinding for the X creator program, I'm finally seeing the downside of the rolling 90-day window. All the early effort I put into testing models like GPT-6 Astra, Sonnet 5.5, and Opus 5.5, Gemini Argon is now starting to drop off. I'll keep pushing toward my y frustrating.

Not so sure about the weekly limit, but Claude's 5-hour limit has definitely been nerfed. On the $100 plan, I now see my usage go up with almost every prompt, no matter how simple or complex the task is. Gone is the excitement of running big tests that barely take a couple of percent off the 5-hour limit. It's not terrible YET, just noticeably worse than a couple of weeks ago. I knew the good times wouldn't last... :(

Another week goes by with Fable 5.5 and Google Carbon getting "released" every single day 😅

Created my own researcher bot with Opus 5.5. My research topic? Male pattern baldness, of course 😅 Burned through $80 of Anthropic's free $100 API credits in ~10 hours. No cure found so far, but here are the most promising leads (there's already plenty of evidence that caffeine owth) 😂 If any AI labs want to collaborate and sponsor some credits to keep this research going until it finds something truly decent, let me know 😉

What I like about it is that it finds the odd, interesting stuff by learning from it's previous research too

Different kind of Opus 5.5 test this time.

US diesel is $6.28/gal. What would happen if…

• Russia replaced Canada as top supplier?
• Russia actually delivered on Trump's diesel deal?
• The Strait of Hormuz reopened?

Getting diesel from russia is barely worth anything to US pening Hormuz.

Data: US Census trade API + EIA. Scenarios: Opus 5.5 estimates.

Opus 5.5 reasoning:

1. Russia replaces Canada (~0¢): imports are only ~4% of US diesel use. Even 50¢/gal cheaper, fully passed on: 4% × 50¢ = 2¢. And a swap adds no new supply.

2. Deal, little arrives (0 to −15¢): Russia's diesel exports fell from 850k to 190k barrels/day this year.

3. Deal, part arrives (−15 to −40¢): diesel futures fell ~23¢ the day it was announced.

4. Deal, all arrives (−40 to −80¢): ~36M barrels covers ¼–⅕ of the world's ~1M b/d shortfall. ⅕–¼ of the $2.80 crisis jump = 56–70¢.

5. Hormuz, 2–3 months (−$1.20 to −1.70): crude back to $75–85 (−35–60¢) + refining margins half back to normal (−$0.85–1.10).

6. Hormuz, 6–12 months (−$2.03 to −2.53): crude back to 2025's $65–75 (−50–75¢) + margins mostly normal (−$1.50–1.80).

I tried the same prompt today to check if there would be any anomalies (nerfing) with Opus 5.5 or if the limits would drain too fast. While I do like the previous result better, I can't say that today's result is much worse. I also didn't notice any big changes in usage, as ed <1% of the weekly limit.

Upcoming previously

I'm trying out a new, fun, and interesting idea with Claude, but can someone finally tell it to stop being my nanny and telling me to go to sleep? FFS! This is about Haiku, not other models, but Haiku still has this annoying habit of reminding me how late it is every freaking minute as soon as the clock hits 00:01.

Me, when I read that "partners" are getting the access to Fable 5.5

Alright, I’ve been hearing people ask if they could try the games I’ve vibe-tested. So, here it is!

It’s not a full website yet - just a domain where I can share the games I make or vibe-test.

You can now try The Pollen Collector and see how it looks and feels. :)
Y

I don't recall laughing so much out of an AI video for a long time, this is just hilarious 😂 Seems like the AI Therapy is still to be solved 😅 youtube.com/watch?v=xTGKP4…

I think I may have reached the bottleneck of Opus 5.5, when I've asked it to create a human 3D model 😂

You can love or hate DeepSeek V4.1 Flash 1017 or whatever, but remind me of a Chinese model that would have been the best model at the time?

I don't know if I am, but I definitely feel shadowbanned. The week where I put in the most of the work is the one of the worst one's. Great.

Yo, did you hear that Pollen Collector V1.1 just dropped? This is what 16% of your weekly Opus 5.5 limit gets you on the $100 plan. Damn. Honestly, this is jaw-dropping to me. And it's not like I sat there all day babysitting it, this was ~16 iterations total. I literally just t out to clean my car, came back to drop another prompt, went back out to clean up the yard, and so on. Can you believe this? Sure, a few things still need some polish (a bit of visual clutter at the bottom during heavy action, the beehive graphics, etc., but overall... Damn!!!). All the technical details and the initial prompt are in the replies.

Stack:• Three.js r170 (WebGL2), the only dependency• Vanilla JS, ES modules via an import map• A 27-line Python static server• 5 free CC0 assets from Poly Haven (1 HDRI sky, 4 PBR textures)Everything else is generated in code.

Every insect, flower, tree and rock is modelled procedurally at load. Insects are skinned rigs: body, compound eyes and shell fur on one skeleton (3 draw calls). Butterfly wings are painted on a canvas and flex in a shader.

Rendering: a custom HDR post-processing chain with ambient occlusion (GTAO), god rays, bloom, depth of field with a macro mode, and ACES tone mapping. Plus cascaded shadows, water reflections, height fog, GPU grass, and a wet look when it rains.

7 maps: Wildflower Meadow, Cottage Garden, Windswept Heath, Alpine Meadow, Lavender Hills, Rainforest Clearing, Rooftop Garden.Day, dusk and night with stars, a moon, fireflies and glowing flowers. Weather: showers, fog, storms with lightning. Plus a photo mode.

6 playable species, each with its own flight feel, ability and passive trait.

Wind with gusts and shelter behind trees and walls: insects fly relative to the air, so light ones get pushed around.5 AI rivals: gatherer, bully, saboteur, explorer, opportunist.

Zero audio files. All sound is synthesized live with the Web Audio API: adaptive generative music, 3D positional effects, rival wing buzz with Doppler, ambience (woods, pond, wind), rain and thunder.

~12h start to finish, ~8.5h of active AI work, up to 4 Claude Code sessions running in parallel. ~16 prompts from me. The AI tested its own work with scripted rounds, frame timing and screenshots.