HN user
Building https://getdot.ai to chat with your data stack.
https://github.com/soofi-project/Soofi-Pretraining https://huggingface.co/Soofi-Project/Soofi-S-Base > "The final model will be released openly under a permissive license, without gated access. We will share access details as soon as it is ready."
not open weight yet
more interesting link: https://arxiv.org/html/2607.09424v2 and > Long-context serving efficiency. Soofi S combines frontier-level capability with the highest measured aggregate long-context decode TPS, and unlike full-attention dense baselines maintains high throughput as context grows. Panel (1(a)) plots Capability Index versus measured aggregate decode TPS/GPU at 40K context and batch 32. The Capability Index averages five benchmark groups, i.e., Code, GSM8K, GPQA-Diamond, English aggregate, and German aggregate, after normalizing each group to the best plotted model. Aggregate decode TPS/GPU is measured with a TP=1, one-B200 vLLM latency-subtraction protocol. Panel (1(b)) shows measured aggregate decode TPS/GPU as a function of input context length under the same batch-32 protocol.
it's a small win in the small model class
Somehow the blog post seems naive. Yes GLM 5.2 is good and cheaper per token, but margins are a result of supply and demand. Now demand for quality and quantity of tokens is increasing at least quadratic or cubic (more users * more tasks * more tokens per task). On the other side you have real infrastructure constraints on the supply side. Openai and Anthropic have large commitments and contracts that enable them to get access at a scale of compute that is not obviously going to be available for open source model hosts. And you see it, glm 5.2 inference is less stable and higher variance than any of the bigs labs.
Why is SpaceX not hosting glm 5.2? because they make more money with renting out to Anthropic and Google.
So he throws billions at a few top AI researchers, but they produce nothing of value.
so he spends 1% of yearly revenue on AI talent to catch up? we can't judge if they have produced nothing of value, no? They don't owe the world to open source their work?
Meta has plenty of failings, but taking risks and investing optimistically is not on my list. I guess the sentiment here on HN is probably biased by the addictive nature of its products.
ASML, SAP, ARM, Spotify, ...
How do you define biggest? Can't be by market cap.
This is called mechanistic interpretability. There is lots of fascinating insights already since you can do basically everything down to the neuron or weight level thousands of times. The human brain is many orders of magnitude harder to make sense of.
We outsourced it for 2.5k (extra) and it was still painful, took almost 2 months and worst of all wasted so much time and focus.
The worst was sitting at the notary, and getting read out loud by her what we were about to sign (also paying for that).
If you think about starting a company, spend some time to think through what it would mean for you to be a Delaware C Corp or an Estoinian one. It will increase your chances of success as you can focus on what matters.
Like sqlite, duckdb is underappreciated as a production database. You can totally run it on servers or even "serverless" and do some heavy data transformations or with the right server size work with large scale datasets (up to a TB compressed seems fine).
I heard the same argument from my doctor when I wanted a blood scan.
But what's the intention? If you do a scan and then try to find everything that is wrong about you, you're 100% right, there will be false positives and unnecessary panic/medication etc.
However if you just collect data for months and years and WHEN you get a symptom you have a lot more data then we should be able to give better diagnosis faster. If we do that for long enough as humanity and there is data sharing the accuracy of the whole thing will increase a lot.
Exactly, like SAP HANA stores everything in memory, you get great analytical and transactional performance but good luck financing that at scale
It was a red herring.
Is LTSE working the way you hoped?
I do think it's more subtle. AI can replace very few jobs end to end with the same quality, maybe none. But AI can be put to work on high ROI problems. Now when the new marginal job is not obviously as high ROI as putting another 100k of tokens to work, no human gets hired.
Next, comes natural attrition in a company where a certain percentage will leave every year. Will they get replaced with a human or their budget goes into tokens?
Only when these 2 angles are exhausted, a typical company will start thinking about layoffs.
Now, some companies are already stressed: customer buy AI products instead of theirs, AI makes it easier to build what they offer, customers believe they can vibe code things. These companies will layoff first, because of AI. Not because AI will do the persons job but because the money gets spend differently.
Don't fall for it. Its a trick question. Capitalism made everybody replaceable a long time ago. :P
The more interesting question is: Has the ease of replacibility increased because of AI?
Naive question but if it works best wouldn't companies that have a four day work week outperform theirs peers and because of that grow faster, and become more common?
I see the opposite in most startups that have a 6 day work week to get ahead of the "slowly moving" 5 day work week competition.
Anthropic will probably IPO as well
yeah 3t/s seems human. only that i never wrote code perfectly top to bottom.
Kudos to the team. Please consider making the model available via API!
In almost every wealthy country the birth rate has fallen below sustainable, meaning we're shrinking if life expectancy doesn't magically explode.
Viruses btw never reflect on killing the host :P
Techno optimist here who expects the following to make a big contribution to reducing human made future climate change: better batteries+solar/wind, nuclear fusion, self driving cars (we'll need to manufacture less cars for the same amount of miles humanity drives), AI helping with better resource allocation in general (hopefully).
The answer can't be, let's just consume 10x less. We have to engineer our way out of it.
Maybe, or they use the same smartphones as everybody else. The mass market also wants the best model and will pay accordingly.
Bubble or not, this is great!
It's priced at 1/10, but deepseek is probably not profitable, also it's slow.
Even more interesting is the question if we would have a deepseek model without the US frontier models.
And then what's the value of the advantage that the frontier models have. It's definitely 100x more valuable to find zero days 3months earlier. Probably not in every domain but in enough domains having the smartest model is valuable.
Down voted because the comment is not providing any new insights and is just insulting.
salaries are opex, data centers are capex, you can't compare them in the same timeframe.
4B over 5 years is 20B, which is significant.
lots of great stuff, but the plot in the paper is just chart crime. different shades of gray for references where sometimes you see 4 models and sometimes 3.
this is vibe coded :D. the drop down for Python not working.
I remember when we started working on our teams bot in 2023, Microsoft announced that they will stop supporting Python for the teams sdk, which felt super short sighted. Eventually they silently picked it up again and the old sdk never stopped working.
Kudos for turning it around. I believe it'll last this time as AI agents in communication tools are something a lot of people want. and unlike pre LLM chatbots and agents, they are so much more useful now.
Naive question but are you not afraid z.ai will train on your personal data?
Hmm, if the stock keeps falling that might really happen.