Pinned Tweet
this is the way.
Replying to @paulg
After many, many years of presuming good faith, I’m finding myself increasingly drawn to a Konmari curation policy: if the prospect of further interaction with a commentator does not spark joy, regardless of whether they are benign intent, block/mute and move on with life.
1
8
6,464
Today one of my clanker's weekly self-improvement loops took more than 5 hours. 🥳
(I should have picked the Highspeed option for kimi-3.1-pro 😐)
(This goes to show how little progress I've made / how early I am in using LLMs and agent harnesses)
1
2
104
If you're using an agent harness you should be able to point it at this thread and at recursive-mode.dev (by @trydotworks) and it should be able to figure out my process. Or just copy and paste the first three posts and the two attached images into it.
1
19
Two hero* trades going into next week to keep an eye on:
1) Japan Tobacco 2914.T
2) Abivax $ABVX
* This is NOT a good thing
If you aren't an existing shareholder you really should be keeping an eye on Japan Tobacco 2914.T
(with apologies to @CacheThatCheque and other existing shareholders)
4
456
To become a net-net, 3690.HK Meituan's share price only needs to fall from HK$67.55 to HK$2.13 (≈97%)
Thanks for reading this useless post, wish you all a nice weekend ahead 😜
2
6
283
But seriously, I've been looking at Meituan all day and, uh, I guess it's all down to the numbers they're reporting next wekk? 😥
1
50
… or not 😰
teapawt.net/AKL_019/status/2094790…
Replying to @AKL_019
7. Given these trends, Meituan’s long-term future is quite clear. One quarter of up or down is not really going to make much of a difference.
44
Previously only available via their PAYG API or the higher tiers of the MiMo Desktop plan, mimo-v2.6-pro-ultraspeed is now also available to MiMo Token Plan Pro and Max subscribers
@XiaomiMiMo
@XiaomiMiMoDevs
1
75
Coocoo retweeted
The last year working with agents has been somewhat painful. We have spent the last few years pretending we enjoy riding iron wheels over cobblestones.
If you look at how LLMs evolved over the past couple of years, the whole industry got obsessed with token efficiency and hillclimbmaxxing i.e. swarm orchestration, RL loops, squeezing every ounce of benchmark performance out of the weights. Everything has been built to top leaderboards or talk to other agents, and all we see in the timeline is news of doom, mass unemployment, and humans slowly diffusing into irrelevance. Somewhere along the way, the actual person sitting at the keyboard has become an afterthought. Living with these models day-to-day now feels like an insane grind. Bookmarking skills/weird prompting rituals, babysitting brittle multi-agent loops, and constantly bracing yourself for the system to either over-engineer a simple question or completely miss what you actually meant (i call this agentic off-by-one error). You spend more energy managing the model's neuroses than thinking about your own problem.
It reminds me a lot of early bicycles. So, in the 1860s, people rode these brutal wood-and-iron contraptions that were literally called Boneshakers, or those absurd Penny-Farthings where you perched five feet in the air hoping a single pebble wouldn't send you flying over the bars. Fast, sure, and impressive engineering, but they rattled your spine to pieces. Plus nobody rode them to relax, the way today you can see kids and adults enjoying them.
And the bicycle didn't really become useful to normal people until John Boyd Dunlop wrapped an air-filled tube around the wheel in 1888, and then the Michelin brothers figured out how to make that tube detachable a few years later. Now adding that air tube was not something insanely difficult and it didn't make the frame any stronger, it just absorbed the shock of the road so your teeth wouldn't rattle out of your skull. IT was born out of dogfooding and care for the person using that machine.
That’s honestly what using the Gemini Neon and Carbon release feels like right now, like a cool breeze cutting through a suffocatingly hot summer day.
After living inside so much jagged, hillclimbmaxxed machinery that clearly wasn't built with human patience in mind, this feels like a team actually dogfooded their own models and decided to put a rubber tube on the wheel. When i play with these models, suddenly this thing doesn't feel like a machine built to replace you, it feels like one built to carry you. You don't have to format your brain into a decryption device of those wild neuralese texts or fight a wall of useless info just to get a straight answer. I haven't prompted five times, explain this to me with a diagram but simpler and use grug like language. I can throw a half-baked thought at it, meander a little, change my mind or style mid-sentence, and the model just absorbs the bumps instead of overly correcting me, getting anxious and pointing fingers at shadows.
To the Gemini team, I will say that this is what we all truly needed and a heartfelt thanks for this. However, unlike past years where something amazing drops once an year and immediate quarters become a repeat spam from openai/anthropic, I wish to continue using Gemini Neon/Carbon successors. I will suffer through this naming scheme, a small price to pay.
Coocoo retweeted
$AUROORA 🇫🇮 Auroora Q3: Organic growth 20,7%, total 36,7%. EBITA +94%. ROCE 20,4%.
Trading at roughly 15 x LTM EBITA.
I've been using qwen4-max for about two days now, from the Alibaba Token Plan (international). Feels just about as smart as, if not smarter than, glm5.3 and kimi3.1-pro.
But the main thing is, it's faster.
I use them from their first-party providers.
1
8
311
Feeling spoiled by the improvements in QoL in @yetone's Alma as well as the rapid update cadence 👍
Took some time to setup @yetone’s Alma with @StepFun_ai step-5-preview (s5p).
I mainly use Hermes Agent and was intrigued after a recent benchmark showed well Alma performed.
Maybe there’s something wrong with my Hermes Agent setup, but s5p does feel snappier on Alma 🤔
1
104
Nvidia card smugglers and illicit buyers, take note 💣
Xbox 360 rigged with explosives detonates at rival repair shop in Brazil, injuring two people
🔗videocardz.com/newz/xbox-360…
What did @LeoPTrudel know?? 😳🤔
Dirty freight rates halved in 2 weeks (!)
USGC to China: 65 M USD to 32 MUSD
Fujairah to Ulsan: WS675 to WS308
WAF to China: WS645 to WS290
This is helping to keep Far East refining margins positive.
#oott
80

















