Game of Thrones, ou encore Astérix et Obélix, voici les dessous des décors de vos films préférés https://www.youtube.com/shorts/TbQ1XH5u5g8 #voyage #culture #nature #decouverte
Search
9 results for “asterix”
-
Looking forward to speaking at #EFNI in Sopot about the future of liberal democracy! 🇪🇺
Europe feels like Asterix’s Gallic village, surrounded by authoritarian powers - with our leaders bending the knee...
We have a magic potion: European unity and integration. But national parties refuse to use it.
We need European political parties like @VoltEuropa to offer what nationalists cannot: cooperation over isolation, leadership over appeasement, and a compelling vision for Europe's future 💜
-
Asterinas 0.19.0 released https://lwn.net/Articles/1100055/
-
Asterinas 0.19.0 Brings Up X.Org & Other Improvements To This Rust-Written OS https://www.phoronix.com/news/Asterinas-0.19-Released
-
Given Kingfisher’s apologia for Neil Gaiman a little while back on BlueSky, I guess we shouldn’t be too surprised by some of the content of the Asterion novel.
HOWEVER, franchise IPs used to be more guarded in terms of quality, in fact I used to read novels set in big franchise universes. They weren’t perfect, and even could be problematic, but the state of publishing was still different in the 1990s.
Hasbro, aka WOTC, should have had at least SOME oversight about the novel, as Mattel should have but didn’t care about the content in an Alex Aster novel, but they didn’t, which is a reflection of an overall publishing trend, and the real or alleged Late Capitalism hypothesis.
https://youtu.be/54RcDKZ8FjI -
Enquanto a classe média compra bilinguismo, o capital internacional compra outra coisa: O grande capital internacional, por sua vez, já identificou algo bem mais valioso neste cenário: a infraestrutura de formação de uma elite globalizada com endereço permanente no Brasil. Esse deslocamento está mudando o setor de modo irreversível https://www.correio24horas.com.br/asteriscao/enquanto-a--classe-media-compra-bilinguismo-o-capital-internacional-compra-outra-coisa-1026 #ClasseMédia #Bilinguismo
-
@bitwarden is now dual-licensed GPLv3 and "Bitwarden License v1.0", which is a proprietary copyright license.
They've gone to great lengths on Reddit, the community forums, and elsewhere to ensure the community that the GPLv3 license remains in place and the clients will remain Free Software and actively developed.
This places them in the same camp as MySQL, Qt, Ghostscript, Asterisk, Cygwin, and others.
But whether or not they stick to their guns remains to be seen.
https://community.bitwarden.com/t/published-version-update-in-app-stores/102750
-
Seven points and one comment. That is what the SubQ Snap launch got on Hacker News, hours in. The comment was a question, and it was not friendly:
Doesn't OpenAI's decision API support up to 1M context on Luna?
On LinkedIn, the same launch hit 220 reactions and 39 comments in fourteen hours. "We launched a new model today!" read one company post. "Feels like Jev on steroids," read an early take being passed around on X.Same product, same day, two crowds reading two different things.
i had to find out which crowd was right. So i read the launch post, the press release, and every third-party breakdown i could find. The short version: both crowds are right, and neither is fully.
SubQ Snap is a decision model from Subquadratic, a research shop that thinks attention is the real bottleneck. No text generation. You send context, a question, and allowed answers; it picks.
The pitch: up to 384K tokens of text and images in one call, a decision in about a second, and five cents per million input tokens. Output tokens are free.
The timing is the wild part. This is the sixth hosted decision model to ship in about three weeks, after Jev, Perplexity's Decider, Cloudflare's Clef, OpenAI's Decisions API, and Microsoft-Decision-1.
Why care? Because this is where the cost of agents gets decided. And the fight is not about the model.
What a decision model actually does
here's a question people always ask: why would you pay for a model that cannot write? A decision model is a narrow cousin of the LLM. Given context, a question, and a set of allowed answers, it returns a probability for each answer.Nothing open ended.
That restriction is the whole trick. No generation means no token-by-token output to pay for. Scoring fixed answers is fast, and speed is the point, because an agent makes these calls constantly: which file matters, which tool to call next, did the test pass.
Subquadratic names four jobs for Snap:
- routing
- tool picking
- checking
- stopping
A clean list. And a list you can currently build with a regular LLM and careful prompts. Which is the part nobody in the press release mentions.
The context wall
what actually happens is the input, not the answer. The right next call might depend on a constraint from forty turns ago.A function defined deep in a codebase. A clause on page nine of a contract.
Standard transformers pay quadratic attention. Double the context, quadruple the compute. That is why decision models stayed small.
Jev runs 64K per request, 32K of state. Microsoft's model runs 32K. They decide without seeing most of what matters.
Snap's answer is Subquadratic Sparse Attention, SSA. Linear cost instead of quadratic. So the model can read a whole repo or contract in one call.
Subquadratic says 100K tokens comes back in under a second. That number is theirs, unverified, no hardware named. But the architecture is genuine research, not a wrapper around someone else's API.
Benchmarks and their asterisks
most tutorials tell you to look at the chart. The chart is good.Decision Index 0.2.1: 61.7 to 57.9. The long-document test: 74.0 to 48.0.
Snap leads on 13 of 16 enterprise tasks. Phishing detection jumps from 25.1 to 47.1. Code output prediction from 57.1 to 73.8.
But read the fine print. Every number is Subquadratic's own. No outside lab has run Snap.
The long-context tests are their adaptations of RULER and LCR. And the headline score, 96.3 on Decision RULER, runs at one million tokens. The product caps at 384K.
They published a benchmark the product cannot accept.
Snap also loses three of sixteen. The embarrassing one is When2Call, which tests whether to call a tool, ask, answer, or decline.
That is one of the four jobs Snap is sold for. It lost 67.0 to 74.6. The model that decides whether to act is not the best at that exact call.
The price war
ModelPrice per M inputContextPerplexity$0.04~262KMicrosoft-Decision-1$0.04232KJev$0.04264KSubQ Snap$0.05384KOpenAI (gpt-6-luna)$0.10n/aCloudflare Clef$0.2464K
the problem isn't the benchmarks at all. It's what the benchmarks are for. Look at the field:
Six models in three weeks. Prices sliding toward zero. Free token grants everywhere; Snap gives the first 10,000 users 100M tokens."Cents per decision" is the new marketing line.
That should make you suspicious in one specific way. When a market moves this fast, the model is not the moat.
The workflow is. The people who win own the routing logic, not a vendor relationship.
Meanwhile, the kidney cats
here's the thing that made me laugh during research. i searched Reddit for "subq snap" and the model was nowhere. i got forums about subcutaneous fluids for cats with kidney failure.i got reviews of SNAP-8 peptide serum for wrinkles. i got a thread about a spring-loaded syringe you snap in place.
"subq" is medical shorthand for subcutaneous. The company named a product after a medical abbreviation and a snack brand's word, then launched it into an AI hype storm. For a while, anyone googling the model gets cat fluids and skincare advice.
i kind of love that. It keeps things honest. The hype lives on LinkedIn, with its 220 reactions and its CEO quote about "radically different ways to scale AI."
The messy internet still serves kidney cats and peptide talk. Every AI launch should have to compete with a cat forum for search traffic. It would keep the marketing humble.
The part nobody wants to hear
Most people do not need a decision model. Not yet. If your agent makes a few hundred calls a day, route with a cheap LLM and move on.The complexity you add to save pennies is complexity you will maintain forever.
Even if you do need one, do not build on a public preview with a waitlist on day one. The model id is literally
subq-snap-preview.That is not a stability promise. That is a warning.
And the benchmarks? Treat them like a resume, not an audit. Everything is self-reported, compared against one rival, on tests the vendor adapted.
The moment an independent leaderboard runs Snap, the numbers will move. i would bet on that.
The one thing that would make me take Snap seriously is a real repo test. An agent that reads its own codebase, picks files, checks results, and hands me the bill. A week of that beats any chart in the launch post.
i'm still going to try it. 100M free tokens is hard to ignore, and that window is genuinely interesting. But i'll treat the ad copy as ad copy until i see it run.
i keep coming back to that Hacker News comment. Seven points, one question, and it cut through the entire launch in a single line. "Doesn't OpenAI's decision API support up to 1M context on Luna?"
That is this market in one sentence. The lead changes so fast that "first" is a memo, not a moat.
Snap is real engineering, SSA is not a wrapper, and 384K in under a second is genuinely interesting. And so was everyone else three weeks ago.
i'll be watching the independent benchmarks when they land. And checking on the kidney cats. One of those two will tell me more about decision models than the press release did.#ai #llm #news #startup #software #coding #development #engineering #inclusive #community
SubQ Snap: The Hype, the Benchmarks, and the Reactions -
CW: NSFW Feral raptor showing off their cloaca and cock
Oh my gosh I got this AMAZING art piece of my raptor form from Vixxie! https://bsky.app/profile/vixxie.bsky.social
They are such a pretty raptor... I wanna snuggle that tail blankie so badddd <333
...but I know something else will keep me warm this night~! ;>