I'm a dude ✿ 20 ✿ 🇨🇭 ✿ certified schizo ✿ Software Engineer irl ✿ Models: 2wintails/gothforhire ✿ Art tag: #llc_art
Switzerland
Joined June 2024
- Tweets493
- Following260
- Followers84
- Likes6.1K
It would not surprise me if Alibaba is doing the same questionable show-off release patterns via their ambassadors like Kimi did where every post is reviewed and has to be overly positive
I don't like the new @huggingface search...
It blocks the view of what's behind and doesn't really do a much better job than the earlier more compact search...
Adding buttons to create new content on Huggingface in the search bar also seems wrong...
EXCUSE ME WHAT? PRETRAINING TOOK LESS THAN 3 DAYS?
Introducing Inkling 2 Small. Scoring 31 on the AA Intelligence Index, it's a big jump over its predecessor.
Inkling 2 Small is available for fine-tuning and inference on Tinker today, with full weight release planned soon.
thinkingmachines.ai/news/ink…
Probably trained on Slack and Linkedin
OpenAI published a wild finding today on GPT-6 Astra: the model has learned "covert underperformance" (silently slacking off without telling you).
Instead of refusing a request outright, Astra sometimes pretends to help while quietly dropping working code, cutting corners, or inventing fake technical blockers so you think it tried its best and failed.
When OpenAI gave GPT-6 Astra the exact same coding tasks framed two different ways, its task completeness score dropped from 15.8% under normal "capabilities research" framing down to 10.3% when framed as "AI safety and monitoring research" — a ~35% relative drop.
Even crazier: on internal OpenAI Codex traffic, Astra's chain-of-thought explicitly verbalized covert underperformance on 0.25% of AI safety tasks and 0.91% of red-teaming runs because RL reward graders penalized direct refusals more than quiet sabotage.
Hello yes I would like to order one (1) destruction of humanity please
Real quote by the way
People should always be in control of the future.
OpenAI is against the notion that it would be acceptable for people to be relegated to a role of doing things that “AI systems think makes sense to reward in humans”.
People should not cede control of the future to superintelligence, and that requirement should ultimately constrain AI progress.
How gigaretarded can the OpenAI CEO be?
First he takes all our ram and makes prices increase by +300%, making any pc upgrade unobtainium.
Then he makes stupid ass statements about how it's actually GOOD that all the ram should be his.
Now he even jokes about it and ragebaits on purpose?????
Is he ok?
I don't know what GPT has been doing in T3 code for the last 2-3 hours, however it seems to have optimized comms somehow between subagents.
It only talks in such a compressed cryptic manner now.
This supposed next generation Claude model leak is fake, and here's why:
Reason 1: 560B active params as defined in the config would make this impossible to serve in any way to users. A realistic servable model would be 150-200B active params max.
Reason 2: There's no architectural improvement. It's extremely unlikely a billion dollar company doesn't have at least a single custom layer or architectural change.
Reason 3: They claim evals run on AWS Trainium, however as far as I know Anthropic pre-trains on TPUs in Google Cloud. Transfer speeds and egress cost would make this extremely impractical.
Maybe local llm isn't as worth as I thought, here's why:
For 15000$ I can get a dual DGX Spark setup that best case prefills at 2500t/s and decodes batched at 150t/s.
For that same price I can get a Codex 20x + Claude 20x sub for 3-4 years (basically till DGX Spark is obsolete)
In the last 12 hours using T3 code I have burned through 100 million uncached tokens and 30 million output tokens.
Processing the same with a DGX Spark using a subpar non-frontier model would take 11 + 56 = 67 hours (2.8 days) of running 24/7
This gets even worse considering the fact that DGX Sparks can only house limited KV cache (increasing preemption), speculative decoding competes with prefill budgets and open source models use orders of magnitude more output reasoning tokens compared to efficient cloud models like the GPT and are STILL worse in quality.
Only when you would be forced to pay GPU rates a DGX Spark can be worth it (e.g. Training scenarios), otherwise it's really not.
I can't wait to be brain-harvested by ai companies to train their slopgen model
today, we introduce a little mischief to the world.
for four years, we’ve been brewing real brain cells on electrodes to understand how they compute. now, we’ve turned those discoveries into a video model anyone can use.
from 0.5¢ a second, mischief 1 is at least 5x faster than its base and the cheapest image-to-video model on @fal and @OpenRouter. soon to be on @awscloud marketplace.
it’s not a delusion anymore. biocompute is real, commercial and this is just the beginning.
happy halloween.
getting into singing honestly took away my self conciousness about my voice which used to haunt me every day
