Tessera

Folgen

Henri Vasserman

@slyecho.net

Dev, Student, Teacher, #NAFO fella, іДі-poster. he/him/his https://est.social/@henri https://mdon.ee/@slyecho

446 Posts1670 Follower 1122 Folgt

Henri Vasserman @slyecho.net · 3d ago

This is pretty heart-warming. youtu.be/AEnJw5G3kNM

2 Kids Just Fixed the Problem Every FPV Pilot Hates

YouTube video by Chase Momo

youtu.be

hailey @hailey.at · 3d ago

this sounds like basically nothing if you read into it. h/t to @vgel.me for a good take on it

H.E. Cas Piancey@cascoinfoundation.org · 3d ago

Is this good?

Cohere: 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 Al 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?

Henri Vasserman @slyecho.net · 3d ago

I feel this

johanna@jola.dev · 3d ago

I'm glad to announce I have finally set up OpenBao I'm sad to announce I now have to migrate all my secrets

Henri Vasserman @slyecho.net · 3d ago

What does this guy have on him?

Social Media Lab@socialmedialab.ca · 4d ago

"Viktor Orbán, U.N. Secretary-General? Trump officials push the idea of a U.S.-Russia-China joint candidate." Yes, you read that right. www.wsj.com/opinion/vikt...

Henri Vasserman @slyecho.net · 5d ago

What if I use nftables

Erica Windisch@ewindisch.ontological.observer · 6d ago

get your iptables filters ready

mr. TIM @timkellogg.me · 6d ago

OpenAI constructed an entire shadow internet for model training so they could remove all internet access they asked agents to try to communicate with each other, and they could: via the CPU cache 🤯 openai.com/index/securi...

prevent. The agents found a way to communicate using the processor's cache: a small, fast memory that holds copies of data that applications have recently accessed.

Henri Vasserman @slyecho.net · 6d ago

help... server? nice

Thought:
Should server help nice. Current env only no flags.

mr. TIM@timkellogg.me · 6d ago

whoah.. uh.. guys?? Kimi K3.1 Pro is eerily close to Opus 5.5 and also open weights and also 3T www.kimi.ai/blog/kimi-k3...

A two-column grid of 14 horizontal bar charts comparing AI model performance across various benchmarks, with Kimi K3.1 Pro highlighted in bright blue:
* Terminal-Bench 4.0: Opus 5.5 leads at 66.4, followed closely by Kimi K3.1 Pro at 65.7, GPT-6 Astra at 58.2, Fable 5.1 at 57.9, DS-V4.1-Flash-1017 at 42.3, and GLM-5.3 at 41.8.
* SWE Marathon: Kimi K3.1 Pro leads at 51.7, followed by Opus 5.5 at 46.3, Fable 5.1 at 45.6, GPT-6 Astra at 42.5, GLM-5.3 at 42.5, and DS-V4.1-Flash-1017 at 40.0.
* Terminal-Bench Science 0.1: GPT-6 Astra leads at 68.1, followed by Opus 5.5 at 63.3, Fable 5.1 at 40.0, Kimi K3.1 Pro at 38.1, DS-V4.1-Flash-1017 at 15.7, and GLM-5.3 at 8.1.
* Kimi Code Bench 3.0 Long-horizon: GPT-6 Astra leads at 66.5, followed by Kimi K3.1 Pro at 61.2, Fable 5.1 at 58.8, Opus 5.5 at 58.5, DS-V4.1-Flash-1017 at 51.5, and GLM-5.3 at 45.5.
* Agents' Last Exam: Kimi K3.1 Pro leads at 43.2, followed by Opus 5.5 at 42.6, DS-V4.1-Flash-1017 at 37.7, GPT-6 Astra at 32.8, and GLM-5.3 at 26.2.
* BenchCAD w/ tools: Opus 5.5 leads at 96.2, followed by GPT-6 Astra at 95.9, Kimi K3.1 Pro at 92.7, Fable 5.1 at 92.6, and DS-V4.1-Flash-1017 at 81.6.
* OfficeQA Pro v2: Kimi K3.1 Pro leads at 64.4, followed by Opus 5.5 at 62.2, Fable 5.1 at 54.5, GPT-6 Astra at 54.4, DS-V4.1-Flash-1017 at 40.4, and GLM-5.3 at 31.1.
* Frontier Finance: Fable 5.1 leads at 60.8, followed by Kimi K3.1 Pro at 58.6, GLM-5.3 at 57.5, Opus 5.5 at 56.7, DS-V4.1-Flash-1017 at 54.5, and GPT-6 Astra at 45.0.
* GDPVal-AA v2.1: Opus 5.5 leads at 1866.0, followed by Kimi K3.1 Pro at 1846.0, Fable 5.1 at 1754.0, DS-V4.1-Flash-1017 at 1746.0, GLM-5.3 at 1650.0, and GPT-6 Astra at 1573.0.
* OSWorld V2.1: Opus 5.5 leads at 81.8, followed by Fable 5.1 at 80.7, Kimi K3.1 Pro at 79.6, GPT-6 Astra at 75.8, and DS-V4.1-Flash-1017 at 69.3.
* Humanity's Last Exam Full w/ tools: Kimi K3.1 Pro leads at 70.0, followed by Opus 5.5 at 67.7, Fable 5.1 at 65.6, GLM-5.3 at 62.5, DS-V4.1-Flash-1017 at 58.8, and GPT-6 As…

Henri Vasserman @slyecho.net · 6d ago

That reminds me to pick up mine from the package machine

philpax@philpax.me · last wk.

💭 I should buy a FM-1

Henri Vasserman @slyecho.net · 6d ago

I had written OPENAPI_KAY on my laptop I didn't see it until I read it on my 5K monitor An AI would not make this kind of mistake

Lum @brokentoys.social · 7d ago

A two-panel comic parodying Disco Elysium using the meme format of You Can See Everything's preview.

In the top panel, Klaasje is drawn with blonde hair and heavy eye makeup, smiling sideways at Harry Du Bois.

Between the panels, a Disco Elysium style dialogue box reads:
KLAASJE (MISS ORANJE DISCO DANCER) - “Why would I deceive
COMPOSURE [Challenging: Failure]

In the bottom panel, Harry Du Bois stares forward with a wide, intensely awkward, toothy grimace, mirroring Nathan Fiedler's uncomfortable expression from the original scene

Henri Vasserman @slyecho.net · last wk.

1m token context. Not seen that on many local models. But it's stretching my 24gigs of VRAM

Adina Yakup@adinayakup.bsky.social · last wk.

Happy to see another on device model 🔥 Spark-X2.6-4B from SparkLLM huggingface.co/collections/... - 4B + Apache 2.0 - Up to 1M context - 3 thinking modes - Local deployment via GGUF

txtxtx @tik-su.bsky.social · last wk.

you know you've made it kui pärast neljatunnist võrgunussi ropendab chatgpt ka vene keeles.

Emilia @lazerwalker.com · last wk.

The 4-panel New Girl / Brooklyn Nine-Nine meme. Jake says “im an encryption expert”. Jess says “oh yeah? Name a private key”. He says “4”, she replies “that’s on me, i set the bar too low”

Henri Vasserman @slyecho.net · last wk.

patch wednesday

What a week, huh?
Captain, it's Wednesday

John David Pressman @jdp.extropian.net · last wk.

In case anybody needs social proof/thinks I'm exaggerating.

A screenshot of 4chan.org showing that the site is down due to the 10/10 PHP CVE dropping tomorrow.

John David Pressman@jdp.extropian.net · last wk.

This is not a tomorrow thing, this is a "if you're reading this take down your PHP based website" thing.

John David Pressman @jdp.extropian.net · last wk.

This is not a tomorrow thing, this is a "if you're reading this take down your PHP based website" thing.

Henri Vasserman @slyecho.net · 2w ago

We have Twitter at home. Twitter at home:

"Xitter" clone of Twitter. Separate backend but same frontend

Henri Vasserman @slyecho.net · 2w ago

So it cares about karma and legacy

Sam Harsimony@harsimony.bsky.social · 2w ago

"We have never seen a model do this before, but GLM-5.3 sometimes reasons (unprompted) about how its current actions could become training data and impact its future self" x.com/ApolloResear... www.apolloresearch.ai/science/upda...

Henri Vasserman @slyecho.net · 2w ago

We have durable agents at home

Q: Is it not possible to separate these events into separate queue items? Instead of one long loop that does tool calls until the final message comes, what if it queues a tool job, then the tool job finishes and queues a call to send the tool output to the model, etc etc?

A: Yes. The turn becomes a state machine stored in the database, with one queue job per step instead of one job for the whole loop. I think it's the right direction, and it fixes more than the timeout. It is a real rewrite of `process-turn.ts`, though, so here is what's involved.

Filippo Valsorda @filippo.abyssdomain.expert · 2w ago

No more leap seconds, ever! There isn’t one in December and they will be abolished in May. The last one will forever be the 2016(?!) one. Didn’t realize having experienced leap seconds in prod dates me. I’m not going to miss time going backwards, I’ll be honest. www.bipm.org/en/cgpm-2026...

Resolution 5 (2026) - BIPM

bipm.org

Henri Vasserman @slyecho.net · 2w ago

I think one thing about Anthropic’s thing is that it’s very nice to have an API visible way of searching and seeing the code without cloning and pulling it. This is even more important when you don’t have a monorepo, because then it’s even more things to sync

mr. TIM@timkellogg.me · 2w ago

Anthropic engineering abandons single global merge queue, suddenly discovering that singular global bottlenecks make scaling hard They’re replacing it with the least obvious solution — merge untested code and hope it’s caught pre-deploy www.anthropic.com/engineering/...

Conceptual Scion workflow comparison. A global-lock merge queue makes changes wait for sequential testing and merging. Scion integrates reviewed changes concurrently, then runs post-merge and pre-deployment tests. Passing tests lead to deployment; failures loop through fixes and re-testing. Scion’s team still reviews every change request, focusing on intent rather than every line.

Henri Vasserman @slyecho.net · 2w ago

False alarm

Xe @xeiaso.net · 2w ago

In case your organization has no soul and hates everything about the word "whimsy": botstopper.net?soulless-cor... I'll probably put the credit card form there in the long run.

BotStopper from Techaro | A smarter front door for the web

BotStopper from Techaro helps protect websites from scraping and automated abuse while keeping the experience straightforward for people.

botstopper.net

Henri Vasserman @slyecho.net · 2w ago

T3 Code is kind of neat

Context handoff from Claude Opus 5.5 to GPT-6.1-Sol

Henri Vasserman @slyecho.net · 2w ago

Sending back the Brother toner cartridges to their recycling center. Feeling like a huge weight has been lifted from the back of my mind

Henri Vasserman @slyecho.net · 3w ago

Me no likey

Felienne Hermans@felienne.bsky.social · 3w ago

If we want to regulate AI, why not start with a simple ban on LLMs using "I"? Easy to build, easy to verify and deflates a hype while still allowing people to use chatbots (so no one can claim we are limiting anything) My latest for @volkskrant.nl in English! www.felienne.nl/2026-10-08/#...

Raphael Wimmer @raphaelwimmer.bsky.social · 3w ago

Reminds me of this comment from 2008 in the thedailywtf.com forums... thedailywtf.com/articles/Cou...

JS (unregistered)2008-01-04
I have a similar experience:

him: Do you have some book on programming?

me: I do. Take this Delphi book. It is fairly simple and should be easy to start with.

him: Ok, thanks. And what language do I do the programming in? (Note: we are not native English speakers)

me: Umm, Delphi. It actually is a mutation of Pascal.

him: Oh no, I don't mean program, I mean language, do I need to program in English or can it be Czech?

me: I am not sure what you mean. Some Pascal keywords resemble English, yes. But you can use whatever language you want for naming variables for example.

him: Keywords? Variables? ...

me: (Explaining what programming actually is)

him: Oh, I thought I would write something like: "Create football stadium and football players. Start the game when user presses spacebar. Make players have red shirts and white socks."

Source:
https://thedailywtf.com/articles/Could-You-Explain-Programming-Please#comment-169043

Henri Vasserman @slyecho.net · 3w ago

Making a system where when agents (or anything, really) use sudo, I get a notification on my phone and I can approve it. Convenient not to type my password on the phone SSH screen. Also convenient to let agents manage the system

Henri Vasserman @slyecho.net · 3w ago

Update on this guy. He even changed his channel name

mr. TIM@timkellogg.me · 5mo ago

the blinding speed at which this guy very publicly speed ran the AI critic to true believer pipeline. Absolutely incredible m.youtube.com/watch?v=-7RD...

Henri Vasserman @slyecho.net · 3w ago

Nüüd tehakse 3 kuu retseptid lisdeksile. Saab osta iga kuu ainult, mitte korraga. Siiski parem, sest ei pea uuendama retsepti iga kuu