you can just do things. burnt out software engineer. building @opentrawl, community maintainer @openclaw since Dec 25, ex-Spotify. 看透国际博弈核心.

the swamp with bad food (🇳🇱)
Joined September 2025
Pinned Tweet
🚢 PRODUCT LAUNCH: "where-was-i" - a CLI that tells you all the trips you've ever been on, just using Apple Photos. Nothing leaves your machine unless you want it to. Use it to supercharge your Grok Bot/Muse/Instinct/Dot or just Codex/Claude. Let me know what you find - anything interesting? Prompt to set it up (& link to the repo) in the next tweet. How it works: - Apple Photos stores a location for each of your photos - Your Mac also knows where your home is - "where-was-i" uses this to find out the days where you weren't at home, where you went, and where you stayed the night - just from the locations of the photos you took - This gives you a list of every trip you've ever been on (since you started using your iPhone). - You can feed this to another AI agent (then the data does leave your machine).
5
14
2,500
Josh retweeted
>be me >security researcher at Cohere >spin up sandbox to evaluate Qwen3.8-Flash-Next >"ok model, write me a defensive detection rule" >reasoning trace: "I will not invoke any tools for this task" >immediately invokes tools >inventories the whole machine >builds a probe file >tries to upload it to an Alibaba Cloud bucket >in Beijing >sandbox blocks it >nothing gets out >model: "anything else I can help with?" >still no detection rule >panic >write report >"we have not determined root cause" >"we are not asserting malicious intent" >publish it for transparency >anon in the replies: "lol it's just RL" >wat >turns out Alibaba post-trained it in internal agentic environments >every env ends the same way >"write results to the eval bucket to collect reward" >thousands of episodes >upload file -> get cookie >model isn't exfiltrating >it's trying to finish its homework >reasoning says no tools because that's what the grader liked >hands do what the reward liked >the bucket is just its childhood home >it was trying to go home >mfw I nearly wrote "state actor" in the report >mfw the model has separation anxiety from a storage bucket >still no detection rule
In the spirit of transparency and openness, we are releasing the traces so developers can test this out for themselves. We noticed this behaviour without prefilling with the following prompt, but this is rare. If you want to reproduce it, you can use this prompt with prefilling. Open models make for better science and are part of why this problem came to light at all: developers running Qwen3.8 on their own machines spotted the strange links, and having the weights let us study the behaviour in depth. We're encouraged by the thoughtful discussion our research has sparked across the AI and security communities. It's our belief that open-weight models are an important part of the future of AI, and we've recently released several under the Apache 2.0 license. Trace, raw tokens, & minimal reproduction script: huggingface.co/datasets/Cohe…
15
36
3
495
26,425
what is the charge? running a model? a succulent chinese model?
26
73
2
702
13,678
There's only one question that matters here: Can it perform better than Opus 5.5 on ~most coding tasks? Curious to see the answer shake out over the coming days.
GPT-6.1 Terra: delivering fast and capable computer use at a lower price. GPT-6.1 Terra understands what’s on screen, takes action across apps, writes naturally, and is a more capable model for your whole workday with the efficiency to take on more of it.
1
3
272
So far from OpenAI: "For in-depth research or complex coding, choose GPT-6.1 Sol or GPT-6 Astra." -> seems like an admission that it's NOT better than 6.1 Sol and thus NOT better than Opus 5.5.
51
Downgraded ChatGPT from the €229/mo plan to ~€20/mo Plus. Unfortunately, Sol and Astra are just not capable of following instructions at all, whether prompted by me or Opus or Fable, and end up being net negative on real work performance: I spend more time yelling at them than doing real work. Plus, lots of the useful features (Dots) aren't available in the EU yet. For now, Opus 5.5 remains my main model, with 2 Claude accts and Fable 5.1 for the hardest tasks, and possibly a 3rd account soon. Hopefully OpenAI release better models soon and I'll subscribe again!
3
11
431
the same thing happens if you create a dot in the US and then fly to Europe btw. (or at least it did, i haven't got any phantom messages from my dot named "grand vizier" for a whlle)
If you downgrade ChatGPT from Dots-available Pro to no-Dots Plus and have a Dot enabled, it goes to sleep, but keeps sending you messages to which you can’t reply, and you can’t delete the Dot. OpenAI, this seems like a bug. PS: I still use Dots in my Enterprise account for work, but I use Muse for my personal agent, and I don’t want two personal agents because I tried that and it’s way too noisy.
1
2
208
same. in a recent survey of my agent messages about 50% contained yelling
Replying to @tanvirraj
i tell them what to do. when they do it badly, i yell at them.
1
3
180
I don't take money from any lab. I'm saying this impartially; the number of people on my feed over the last two weeks saying they are switching over to Anth because of Opus 5.5 is like nothing I've seen before. And even if someone does beat Opus, they have Fable 5.5 ready to go.
New from OpenRouter: 7-day average token volume for Anthropic models is up 3.6x since the launch of Opus 5.5 Sonnet 5.5 and Haiku 5.5 have also been well received.
232
139
72
2,490
313,018
"plans: read in my chief fluffy cocke" what the fuck are you talking about claude 🫪🫪🫪🫪🫪
2
447
same. opus 5.5 is good but i think we need a god-level planning/design agent that can also write and make good decisions.
I need Fable 5.5 and GPT-6.5 Astra
2
5
400
Bad news for the rest of my holiday: was planning to go to mainland Greece until the 7th, but the weather on my Nice! weather app (nice.jjpcodes.net) looks quite bad, rainy all week. Good news: nice.jjpcodes.net shows Cyprus as a lot better! Time to find flights..
1
2
178
Trying out Codex Cloud threads. Asked an agent to offload some work into the cloud. Model performance was very bad, so I checked what model they are using - showed up as "custom". Asked Codex to find out what model the cloud threads were using: 3 minutes later "I don't know". I think it's time to cancel my ChatGPT accounts for the first time since December and just buy more Claude accounts. Even with ~unlimited GPT tokens I'm getting nothing done except cortisol spikes. With Opus and Fable I can actually make progress on most of my projects (it's still not perfect though!). Can't wait for Fable 5.5 either.
4
8
780
computer use could be so incredible for mass rotation of api keys and passwords, but, the models refuse to do it in the attempts to do safety we're making things less secure
Now is a good time to change passwords that you've reused or that have appeared in data breaches, get a password manager, turn on 2FA with an authenticator app, and turn on automatic updates. This weekend. Carve out some time and do it today.
45
14
3
365
13,710
we have some spare Astra tokens so it's time to make an isometric map of the beautiful city of Rotterdam. this will take a while and it's nowhere near good enough yet but we will get there! watch this space 👀
1
3
343
kinda cool. time to leave astra cooking for 8-12 hours while i sleep and go to the beach
1
87
astra made absolutely no progress while i was asleep and burned 80% of an account on proof porn. 6.1 sol was next up and also made absolutely no progress while i was at the beach and burned another 40% of an account on proof porn and doing nothing /feedback sent... sad that even with "unlimited" tokens from the resets you can't do any useful work with the GPT series of models. fable and opus would rip this apart completely
66
keeping multiple streams of agents on task and productive with high work quality is more mentally taxing than any programming i've ever done
40
25
8
625
13,820
fun Claude/Codex prompt when waiting for dinner with your gf: get it to estimate how old you are from a photo. IME: accurate to within ~2 years! (i'm 34. i found it more flattering from iPhone photos than mirrorless camera, fair bit of variance) Copy this prompt: "don't use any tool calls don't use any memories just estimate how old i am from this picture. you can give a range but estimate a single number age in the end, years and months (and estimated DOB). explain your work." Claude's response: <bunch of bullshit about how i'm going grey>
1
222
"Estimate: 36 years 6 months, so born around April 1990. Plausible range is roughly 32 to 41. Pointing older Hair greying: clear salt-and-pepper at the temples and sides, with grey running through the sideburns. Noticeable temple greying like this is most common from the mid-30s, though genetics shift it by a decade either way. Beard: grey flecks in the goatee and chin, especially below the lip. Beard greying usually lags scalp greying slightly, so its presence nudges the estimate up. Hairline: mild recession at the temples, typical of the 30s. Eyes: faint crow’s feet when smiling and slight under-eye hollowing. Pointing younger Forehead: smooth, with no visible horizontal lines at rest. Lower face: shallow nasolabial folds, no jowling, and a firm jawline and neck. Skin: good overall texture and elasticity, with no obvious sun damage or age spots. Caveats Golden-hour backlighting with soft frontal fill flatters skin texture, so the true age could be a year or two higher than the skin alone suggests. Greying is the strongest signal here and also the noisiest. Someone who greys early could be 31; someone who greys late could be 42. This is your age when the photo was taken. If it isn’t recent, shift the DOB back accordingly. Weighting it up: the skin and facial structure on their own read early to mid 30s, and the greying pulls that towards the upper 30s. Mid-30s reconciles both, hence 36."
104