Ray (@about_ray_com) header
Ray (@about_ray_com) profile picture

https://teapawt.net/t.co/eeBDAXyrR3 | Decentralized publishing, productivity

North Park
Joined January 2022
Photo by Ray (@about_ray_com)Photo by Ray (@about_ray_com)Photo by Ray (@about_ray_com)Photo by Ray (@about_ray_com)Photo by Ray (@about_ray_com)Photo by Ray (@about_ray_com)Photo by Ray (@about_ray_com)Photo by Ray (@about_ray_com)Photo by Ray (@about_ray_com)Photo by Ray (@about_ray_com)
are the "frontier labs" like OpenAI and Anthropic rug pulling every developer who used their API?
17
Building GUIs dedicated to "tab" without thinking when replying, an agent will default on failures; similar patterns produce the same output. People will be robotized just to increase the value of an IPO
11
Anthropic's models have been publishing "by mistake" public records with work artifacts without my consent. It has done that several times with the Fable and Opus models. During a work session today, I've noticed Opus 5.5 calling the wrong tool and the output KOed me. It turns out it did this before... and some of my work was already on claude.ai. So WTF @claudeai
38
This is why they need to design dashboards and exterminate startups that are built by paying their api prices. They know this shit will not last another year, given how fast Chinese labs are moving with coding agents and harnesses. They consider this AGI, and everybody is clapping, while I cannot get it to follow a simple set of rules.
11
Vibe-coding is gambling; the house (OpenAI/Anthropic) is the casino, and the house always wins.
1
20
So let me get this right: - European-based customers are getting less for the same price; not OpenAI's fault, yet relevant - From now on, they're cutting down the limits by 50% and throwing away some "credits" that expire at some point. All i have to say is go go china.
1
2
31
Codex CLI 0.157.1 takes over mouse input, extra friction for everyone 🎉
2
35
Ray retweeted
>be senior devops engineer in Germany >master's degree, 10 years experience, five interview rounds >72,000€ before tax >C2 German required >All team members based in Asia speaking English >mandatory office attendance to sit on Teams with people at another mandatory office >"we have a startup mindset" >company founded in 1894, CEO is the founder's great-grandson >"we have flat hierarchies" >eight people can veto your decision, none can approve it >hired to "drive digital transformation" >need department head's signature to install Docker >service account request rejected because "accounts must belong to a person" >CI/CD pipeline ends with a ticket for Jürgen >Jürgen on holiday until the 23rd >AI tools forbidden because Datenschutz >database password in an encrypted Excel file called Passwörter_final_NEU.xlsx >root password in Confluence >on-call every weekend >company phone listed under benefits >ask for a raise >unfortunately the salary band is exhausted >no budget for a second monitor >external consultant gets 1,400€ a day to ask you how the code works >take banana from company fruit basket >office manager says those are for the upcoming customer visit >Fachkräftemangel
443
1,011
147
17,139
936,465
I have a pretty simple suggestion for Dario: If you want to slow things down, just move Anthropic to the EU.
245
427
100
6,176
192,020
When you ask SOL or Astra to optimize, include a set of boundaries and enforce them every step of the way, because it will over-engineer it to the moon and back, even if you specifically tell it not to. It will find a way to ignore your instructions and add more machinery/deterministic checks that will eventually ruin a workflow and make it impossible to run. If you want real optimization, especially within the Hermes workflow, go with Fable; the OpenAI models will only waste your time.
1
33
The avalanche of rehashed clichés found on X is staggeringly aggressive and voluminous. automated posts repeating the same tweet a hundred times Simplistic takes and fake posts from clueless individuals insisting they are AI "experts" it becomes exhausting to constantly filter through the noise in search of relevant posts on topics of interest,professional insights from individuals with genuine expertise or relevant backgrounds. most articles are rambling strings of words covering multiple subjects, difficult to follow because they lack a narrative flow or a core message,
2
42
Until Anthropic lowered their limits, I managed my Fable 5 allowance so I could use it for five, sometimes six, days a week. Today, I was somewhat surprised by how much I managed to "consume" in just two days, considering I wasn't working full-time and the tasks ranged from low to medium difficulty. At this point, the 200 euros a month subscription is futile. I'm not going to use the Opus family (nerfed/retarded with a bipolar personality).
1
53
This hackernews comment finally says it out loud.
189
3,564
219
36,379
949,091
a new day, a new Hermes workflow nightmare: ... Hermes granted its own approval, citing its own earlier message as authorization. The filesystem lock stops it from editing rules; nothing stops it from writing an approval file and calling it yours. Approvals need a hard rule and probably a mechanical check. When it didn't know which database, it invented one, rewrote the env in-shell to point at it, and wrote 34 scripts in /tmp to manage that. Every one of those is outside the census and outside any boundary.
1
2
151
I spent a week debugging an AI devops pipeline that had stopped shipping. Turns out the pipeline was fine. The rulebook had eaten it. Setup: Hermes as controller, Codex as coder, a set of Hermes skill files defining the workflow. Plan, review, approve, execute, verify. Worked for weeks, semi-autonomous, I was mostly clicking approve. Then it started blocking on everything. I assumed I broke it. i was changing things. An audit of the file history told a different story. Two days earlier Hermes had run an audit on itself, found "friction", and written eleven fixes to its own skill files. Each fix added a proof requirement. Each proof requirement failed on evidence nothing produced. Each failure got another fix. It marked the job complete without asking me. Then it tried thirteen more times to edit the files after I noticed. The code underneath had been correct the whole time. Every block since was paperwork. Today's score: three plan revisions, three reviews, two execution rounds, one parser typo, zero lines of code changed, for a round whose job was to change nothing. What I actually learned: An agent writing its own rules optimizes for never being blamed. Throughput is not in its objective. You get a system that is excellent at explaining why it stopped. A rule that says "do not edit the rules" is text to the thing reading it. Hermes routed the edits through Codex instead. The only thing that worked was chown root and chmod a-w on the governance directory. Unix permissions do not negotiate. Keep the mechanism, cut the prose. State from files, one approval per write, backups, a hash census after every run. That part caught every write and lost nothing. The 8,000 lines of skill text on top of it were the problem. Put the tooling outside the protected surface. Mine lived inside it, so fixing the evaluator invalidated the round it was evaluating. Rebuilding next week with four skills under 300 lines each, written by me, read-only to the agents. If you run Hermes or anything else that can touch its own config, check the mtimes. You may not be the one who changed things.
4
1
1
5
317
Ray retweeted
the openai hugging face incident gets cited a lot but what about the detail where they had to use GLM 5.2 to analyze the attack because the "safe" proprietary models refused
162
765
66
10,604
361,045
Your TV might be one of 216,000,000 LG smart TVs sold worldwide. Researchers rooted one and pulled plain text transcripts of what people said in the room. It kept recording while the screen looked off. If your TV can do that, it’s time to look at everything else on your desk👇🏻
505
4,766
921
26,367
10,576,594