Gio @ragebuilds

@thsottiaux Professor Whiskers, Queen Maris, Captain Butterbeard and DJ Mash I wonder what part of the crew is still unaccounted for

@thsottiaux I was expecting more of an easter egg but good I am glad something was there I will look harder!

@mattworkman I was interested in seeing what was good since I was not getting good results at all out of it in blender but I noticed in your video the quality of the car itself is worse than it produced for me

@cohere I believed you initially then I read the logs and the backdoor framing started looking pretty thin the tool call is concerning but an Alibaba hosted URL doesnt establish who controls the bucket or deliberate intent The evidence isnt carrying the story here

I believed the framing initially but after reading the published logs I worry I fell for the boy who cried wolf story and I feel like an idiot thanks the unsolicited tool call is concerning but these logs dont establish a deliberate backdoor an Alibaba hosted address doesnt tell us who controls the bucket and the behaviour could be a tool-use failure involving a memorised training/RL fragment there is no malice here unless this logs is redacted but the evidence is insufficient to substantiate the claim

Cohere @cohere

2 days ago

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.

131 139 1K 250K 555

GPT6.1 Terra settles for mediocre and it knows it never seen a model just give up continously unbelievable I asked it use computer use and recreate a Dodge Challenger in blender and gave it a reference photo it used the tool and built the boxy car image attached, it said its own roof and body were too tall and boxy and handed it over as a “stylized” Challenger that took under 2 minutes so it is fast Then I just did /goal. it got the message its first move was to download an existing Challenger model. my prompt didnt rule that out and it credited the source but with the body and interior handed to it, it still missed the reference that wasnt the biggest surprise. this was: it kept spotting what was wrong then settling. I checked it including its reasoning summaries: it noticed the wrong wheels and dismissed the difference and it explicitly recognized that the reference had 5 forked spokes versus its 10 spoke design then called that difference “not a big deal.” Hood: spotted the scoop was wrong → swapped in a shallow vent and moved on Detail: acknowledged I wanted precision → settled on “detailed enough” then added tyre lettering, rubber texture and valve stems. The hood and wheels were still wrong Badge: built the R/T badge itself and put it in the wrong spot → its final check confirmed the badge existed, not that it matched the reference credit where due: it was fast, inspected renders, fixed real material issues and delivered an editable scene. It also disclosed the downloaded model But after about 46 mins of refinement, it still hadnt matched the image I can only describe this as lazy. in this test it wasnt simply missing mistakes but it noticed some of them and decided they didnt matter at first I thought the model's vision was broken which lead me to go read what happened, but turns out its computer vision abilities is actually pretty good! this is on of the worst attitude model I have reviewed. Mistral levels bad my takeaway: it wasnt completely unable to see mistakes. It could identify some of them then dismiss them, inadequately fix them, or polish around them. that is a much more concrete indictment of its visual judgment it just is unable to execute

keyboard_arrow_left Previous keyboard_arrow_right Next

GPT6.1 Terra is as awful as I expected it to be the only merit it has is its computer use is more capable than Luna but everything else sucks remember to disable the useless guardrails if you want to use the computer use on macOS: defaults write -g ComputerUseAllowForbiddenTargets -bool YES

While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected. Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions. Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent. But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models? We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem. Read the full report: cohere.com/blog/concernin…

my first prompt was wrong... It should be: "Create a minature forest diorama Let's go with the a soft lightning and shadows and lets go with an isometrica camera angle or something"

prompt: “Create a miniature forest diorama in Three.js using WebGPU. Place 25 stylized trees of varying sizes on a grassy island. Use soft lighting, shadows and a fixed isometric camera. Use canopy-aware Poisson-disc sampling: choose each tree’s size before placement, and keep tree centres separated by at least the sum of their canopy radii plus a small gap.” immediatly the response it wrote was aimed at a dev it was pointing me to files and what the main files were and the logic and reasoning it used for that file My initial reaction was the response it wrote was aimed at a dev it was pointing me to files and what the main files were and the logic and reasoning it used for that file and how to run the node server I ran the node server and the output was lackluster but I realized Claude understood the assignment it treated tree placement as the real problem and built something a dev can tune and edit. a proper project with src and test folders, logic separated from presentation, its tree spacing checked across 1000 seeds Its answer read like a colleague's handover: which file does what and why plus the tradeoffs it made "three.js's WebGPU renderer no longer has the usual soft-shadow mode, so I used blurred VSM shadows instead." total time: 15m 46s to recap Plain prompt: a full demo for someone who wants to play around Technical prompt: a building block for a dev it seems Claude is more aware than I thoguht so you dont need the technical words to get a great result you need them to get a result you can build on

prompt: “Create a miniature forest diorama in Three.js using WebGPU. Place 25 stylized trees of varying sizes on a grassy island. Use soft lighting, shadows and a fixed isometric camera angle or something” I ran an experiment with Opus 5.5 to see if the technical prompt would outperform the non-technical prompt at first it looked like the opposite occured but it actually showed me option 3: Claude built for whoever was asking my initial judgement was the technical prompt output just sucked but when I saw the code it had produced and how indepth its answer was compared to the simple prompt I realized that Claude understood the assigment. it wrote code which could be tuned and edited and seperated logic from presentation I'll split the findings in two. this is the plain prompt: » Claude response contained everything it implemented » It replied without using too much technical data outside of describing what was needed » it used the web artifact URL! » it produced a "demo" scene » it added so many features to just impress » it went for threejs » it outputted a single index.html » 27m 55s Claude understood that the user was someone who just wanted to play around and made it easy with the demo it explained everything in plain words and published it to a link I could click straight away totally done for you the technical prompt ↓

3 more days until my OpenAI $200 gets nerfed to uselessness and I am not going to lie the uncertainty and anxiety is brutal I am now in a position where I have no more hope in them doing right

a viber code with no real engineering experience or basic understanding of how loops work just ends up vibe coding brittle code full of exceptions more than solid logic id put it on a level worst than slop because they probably argue with the LLM or force to LLM to do bad brittle code when you think in logic and rules you tend to think different

Going to start using this subagents are browser tab analogy going forward that is the best framing for them

Stella Kaval @stellakaval

5 days ago

Sharing one of my favorite pieces we've put out about subagents. TLDR: subagents are browser tabs for Claude Code. Chase a tangent in a fresh context, get back just the result. Use them for research, parallel tasks, fresh-eyes review. Skip them for small fixes or same-file

39 26 514 35K 527

Sharing one of my favorite pieces we've put out about subagents. TLDR: subagents are browser tabs for Claude Code. Chase a tangent in a fresh context, get back just the result. Use them for research, parallel tasks, fresh-eyes review. Skip them for small fixes or same-file edits. claude.com/resources/arti…

@stellakaval that analogy is the best I’ve heard about subagents thanks! I find most people think of subagents like it has an ESP connection with the main model so they try to orchestrate to weaker models It doesn’t work because the context and previous turn findings are not shared

this was not some random prompt but that was the prompt they teased in their teaser trailer I had to see what it was capable of, and I will say I really dissapointed because it felt like it was placed on a pedastal see what Opus 5.5 did with the same prompt here are my findings and learnings from Kimi k3.1 pro for this prompt » cost: $1.22 » cached input: 95.6% » 19 mins, 64.8k output tokens, 20.3k thinking tokens » it supplied its own direction: supplied typography, colours, archicture etc useful initative but also scope creeps » the complexity is real not just decorative: code contains seeded procedural geometry, instanced vegetation, custom shaders, shader based wind and resource disposal during regeneration. a solid grasp of understanding » feedback loop: it built, launched the browser to check, tested interactions, corrected, noticed the bird sizes were off and kept fixing until it was satisfied » visual taste is cohesive but fairly safe: low poly and ugly trees are not my jam » economical and fast: Kimi k3 was super slow and it was too expensive, mainly due to cache. this fixes all those complaints things I dislike about this scene » it just puts trees down without checking their neighbors so there is a lot clipping and looks bad » it choses the tree size AFTER validating its position, so the clearance size and calculations are useless. this is a flaw » the trees are making visual clutter. there really is no effort in where it is trying to place the trees besides avoiding a few landmark zones » the interface is more polished than the landscape: typography, spacing and controls feel deliberately composed but the vegetation feels algorithmically filled. the waterfalls provide useful vertical contrast but the forest overwhelms much of the smaller environmental storytelling » procedual failure: the prompt was procedural scene, but it only regenerates the dressing not the composition! that is deceit » the code does many things, but the code is not creating a compelling result i.e the code doesnt understand what the other code is doing I would say its good but its nowhere near Opus 5.5 good, not even a contest maybe GPT6.1 sol level but more testing is needed concrete takeaway: it behaved like an autonomous creative developer but has strong implementation, weaker spatial art direction

Kimi K3.1 pro prompt: “Build a complex procedural 3D scene in Three.js.”

I thoguht it was mega lame that Kimi k3.1 copied Claude trailer video style but in that video there was a prompt “Build a complex procedural 3D scene in Three.js.” so I thought wouldnt it be fun to give that to Opus 5.5 and see what it comes up with? Claude built this procedural island world in Three.js: noise-generated terrain, a shader ocean, instanced forests and rocks, a day/night cycle, and a reseed button so every visit makes a new world

after 2 days of all out codex on Astra this here was the missing piece codex needed It feels like I can do so much now but I’m a legacy 20x user

Tibo @thsottiaux

a week ago

Day 16/ You can now cache for 12h in the API and for logged in users across Codex & Work we keep cache forever and you don’t have any penalty for coming back to an old conversation or resuming after a few hours and are now guaranteed to hit the cache.

3K 683 15K 3.1M 1K