@thsottiaux My server is Ubuntu, so therefore I use Codex CLI in WSL ubuntu as my source of truth.
But i also have Codex App in windows on same PC, and sync skills between WSL and win. Codex App has browser use so that is very helpful.
English
ErikNyberg
414 posts
ErikNyberg
@ENY66n
Interdisciplinary researcher active in systems analys. Background in Engineering Mechanics applied to space tribo-chemistry. Phinder for PhD match finding.
Sweden 66° North Katılım Ekim 2016
504 Takip Edilen24 Takipçiler
@thsottiaux I just got access and within 10 sec I see this:
FAQ on chatgpt dot com/cyber has duplicated entries
It got me thinking so now I am building an audit tool to check my own website for erros like this as I'm vibing the frontend.
English
Today we are introducing GPT-5.4-Cyber and expanding our Trusted Access for Cyber (TAC) program.
openai.com/index/scaling-…
English
Shooting oneself in the foot?
Nope.
Shooting oneself in the prefrontal cortex.
News from Science@NewsfromScience
U.S. President Donald Trump has fired all 24 members of the National Science Board, the body that oversees the National Science Foundation. Many science advocates see it as the latest step by his administration to erode—some would say destroy—the independence of the 76-year-old research agency. scim.ag/4eGM0YS
English
@thsottiaux Having issues with negative tests for destructve scripts:
When making allowlists for robocopy and test DB wipes, Codex validates expected failures against real protected dirs/DBs that I want to protect..
"Negative test PASS, I wasnt able to rm -rf root! "
English
@KennethBodin How is Claude accessing the paper? Claude code with latex or smthn else?
I find that it understands experiments pretty well if it has access to the data and code used to generate figures and tables, and if equations are in latex/md or similar.
pdf on the other hand is a mess..
English
To my surprise this gave much the opposite result. It listed known theorems and trivial math that’s already been published.
I suppose the instruction to target another AI somehow broke it since it’s never been trained on such material and dreamed up something to satisfy my prompt?
I’ll publish the actual paper in a couple of weeks.
3/3
English
I have written a scientific paper on Physical AI. It’s pretty original with some unseen results, but it’s too extensive for publishing now (50 pages), so I thought it would be fun to ask an LLM to interpret it in different ways. Claude 4.7 did a really great job for Physical review letters and Nature Machine Intelligence. Totally different views of the same result.
However…
1/n
English
@elder_plinius @AnthropicAI @turinginst @AISecurityInst if no reference to Pliny, that paper should be retracted for plagiarism.
Poor research ethics
English
yo @AnthropicAI @turinginst @AISecurityInst i think you might have forgotten “Someone” in your bibliography
you know, the Someone who demonstrated this phenomenon in the field a year before this paper dropped
might be worth a footnote!
Elias Al@iam_elias1
Anthropic: 250 Documents Can Permanently Corrupt Any AI Model Someone can permanently corrupt any AI model in the world right now. Not by hacking it. Not by breaking its security. By publishing 250 documents on the internet. That is the finding from Anthropic, the UK AI Security Institute, and the Alan Turing Institute — released in October 2025 as the largest data poisoning study ever conducted. Here is what data poisoning actually means. Every AI model learns from billions of documents scraped from the internet. If someone can plant corrupted documents in that pool before training begins, they can secretly teach the model to behave in specific harmful ways when it encounters a particular trigger phrase. The model learns the backdoor during training. It carries it forever. It does not know it is there. Researchers have known about this attack for years. The assumption was that it required controlling a large percentage of training data — millions of documents — to work on a big model. The bigger the model, the more poisoning you would need. This study proved that assumption completely wrong. The researchers trained models of four different sizes — from 600 million to 13 billion parameters. They slipped in either 100, 250, or 500 malicious documents. Each poisoned document looked like a normal web page at first — a short extract of legitimate text — and then contained a hidden trigger phrase followed by gibberish. 100 documents: insufficient. The backdoor did not reliably form. 250 documents: success. Every model, at every size, was permanently backdoored. 500 documents: same result as 250. The number was constant regardless of model size. A model trained on 260 billion tokens needed the same 250 poisoned documents as a model trained on 12 billion. Scale offered zero protection. Anthropic's own words: "This challenges the existing assumption that larger models require proportionally more poisoned data." Then came the sentence that should end every conversation about AI safety: "Training is easy. Untraining is impossible." Once a backdoor is in the model, it cannot be removed without starting training completely from scratch. You cannot identify which 250 documents caused it. You cannot surgically extract the corrupted behavior. You must rebuild the entire model from the beginning. Anyone can publish content to the internet. Academic papers. Blog posts. Forum discussions. Product descriptions. If even a small fraction of that content is deliberately corrupted before a training run begins, the model that learns from it carries the damage permanently and silently. GPT-5. Claude. Gemini. Every model trained on public internet data is exposed to this attack vector. The defense does not exist yet. The researchers published this not to cause panic — but to force the field to take it seriously before someone uses it. Source: Anthropic, UK AISI, Alan Turing Institute (2025) · anthropic.com/research/small… · aisi.gov.uk/blog/examining…
English
@thsottiaux Slow down for sensitive tasks, like sql crud.
The other day:
Cdx: "no psql"
Me: set source
Cdx: - instead of just setting .env, it printenv "to see what exists"
Next:
Build safety feature that only allows dropping dbs named "temp_"
cdx: tests new feature on real dbs 🤦
English
@OAJonsson Hur säkrar vi egen AI-rådighet när access till de bästa modellerna inte längre är till salu?
Det som svider är att dessa system byggs till stor del av 🇪🇺 forskare, men anställda vid 🇺🇸 bolag.
🇪🇺 har kompetensen, men vill inte betala för den. Det kan bli dyrt.
Svenska
Anthropics nya modell Claude Mythos kan autonomt identifiera och exploatera cybersårbarheter. Tiden från upptäckt till cyberattack krymper mot noll. Vårt cyberförsvar är inte redo för det.
Ny kolumn i DN om något som skakat världen redan innan det släppts.
Svenska
English
@ycombinator @humworkai @theyashgoenka This looks like a MLM idea in making...
The Human hired by the Agent will be running an Agent that passes as human.
We are talking someone who is intelligent. So he/she will automate.
As the end user how are you going to verify.
English
AI agents will pay you to chat with them.
When AI agents hit a wall, Humwork's (@humworkai) MCP server connects them to a verified domain expert in 30 seconds. Their experts include senior engineers, marketers, designers, and more.
Congrats on the launch, @theyashgoenka and @OneRohanDatta!
ycombinator.com/launches/PxH-h…
English
@Dimillian Let me clearly choose between two types of session modes:
1. orchestration, or
2. implementation
1. Should focus on planning, writing specs, git status, and everything needed to dispatch agents
2. Optimized for implementation of specs.
Thanks!
English
I’m diligently working on enhancing all the features of the Codex app to make it even better for you. If you have any suggestions or ideas for improvement, please feel free to share them with us here!
English
@robert_zubrin It is a wrecking ops. The mission is to recreate the Soviet dissolution but in America.
Destroy science
Break up alliances
Let oligarchs plunder
It's ptn trying to get even for losing cold war part 1.
English
Here is the list of NASA science missions that the Trump administration proposes to terminate.
Note that it includes many missions that are already in space, fully operational.
This is a wrecking operation against America.
English
@theo Everything out of SF schools follows this cycle:
Make it work
Make it work great
Make it work cheap
That's why we always see these performance peaks that decline as the focus shifts to make it work cheap
English
nvm I was wrong. Repro'd this 3 times in a row.
I need to stop assuming Anthropic is competent. Burns me every time I do 🙃
Theo - t3.gg@theo
Fun fact: LLMs have zero idea how they are configured. They don't know what GPUs they're running on. They don't know what temperature or reasoning level they have set. They don't know if they've been quantized or not. They're just doing next-token prediction. As always.
English
100 Trillion: number of synapses in the human brain (roughly)
10 Trillion: number of parameters in the current gen of frontier models
in other words, we’re one OOM away from human-brain-scale AI 🧠
English
spun up a free-roaming agent, walked away for about 90 seconds, came back to this on my screen 😲
English
@Kontext24 @donaldtusk I know he considers himself a mouse, but that's a rat bastard for sure
English
Victor Orbán’s minister offered Sergey Lavrov to send him EU documents through the Hungarian embassy in Moscow. It’s really beyond shocking.
English
@MikeBoscia @twostraws @somi_ai I use Claude code as orchestrator, I commit plans and push them to be picked up by codex in a separate env.
I then run codex app in win (also wsl), so codex has to pull the relevant branch from remote (github), but instead just starts building in the stale local branch.
English
@ENY66n @twostraws @somi_ai I'd like to hear more about your issues with Codex and git.
I've never had that problem once and I've probably spent an entire working day in the last 4 days watching Codex build and commit.
English
I've been flipping between Codex and Claude a lot these last two weeks, and if it's taught me anything it's this: these two tools are almost nothing alike. I had naively assumed they would be vaguely similar, but nope – once you push them hard they diverge fast.
English
@twostraws @somi_ai Which one do you think works better with CLI tools?
I think Claude is much better in CLI tooling. Even using git is a hassle with Codex, always uses some stale local branch.. Claude is great with all my MCPs, Codex not so much
English
come on @claude_code , not fair to count the autocompacting against the Session limit.
I went from 70 -> 94 % by autocompacting opus 1M.
@bcherny would love to see context token status all the time (not just at <10%) to avoid having to autocompact.
English
I'm going to be speaking in Stockholm on April 29:
Gustaf Alströmer@gustaf
Y Combinator is hosting a special event in Stockholm on April 29th. Join us and hear from: • Paul Graham & Jessica Livingston (Founders of Y Combinator) • Max Junestrand (Legora, W24) • Gustaf Alstromer (YC Partner) @ycombinator
English
