Public sector
In Sarawak, we helped fine-tune an LLM on an indigenous language using federated learning, then run it with distributed inference.
Most government data can’t leave its source. With federated learning, it doesn’t have to.
https://www.flock.io/blog/flockio-is-the-key-technical-partner-in-a-sovereign-ai-government-initiative-for-sarawak
Reply to @flock_io: General-purpose models know the internet. They don’t know your claims process, your case law or your patient pathways.
Most of them also run in someone else’s cloud, so every prompt sends your business data outside your walls.
A model fine-tuned on your own data gets the terminology right, answers faster, and has a smaller footprint. It costs less, and it can run on infrastructure you control.
35/100·CLong
m
meme10/25meme
@pukerrainbrow Throughput changes what agents are for. A 10-minute task gets queued and checked later. A 20-second one you iterate on live, retry and refine while it's still in your head. Same model, different product.
Reply to @pukerrainbrow: An AI agent task that takes 10 minutes at 100 tokens per second would take about 20 seconds at 3,000.
Gimlet Labs wants to deliver the second number. It raised $300 million at a $3 billion valuation less than a year after leaving stealth.
The company is based in San Francisco and led by CEO Zain Asgar, a former NVIDIA GPU architect who also led engineering at Google AI and teaches as an adjunct professor at Stanford. His co-founders include Michelle Nguyen, Omid Azizi, and Natalie Serrino. The four built Pixie Labs, a Kubernetes observability startup that New Relic acquired in 2020.
The problem, Gimlet says, is that many inference clouds are built on one kind of GPU, which forces a tradeoff between total throughput and the speed each user experiences. It argues that inference isn't one uniform workload. Processing the prompt, generating the answer, attention, and feed-forward layers all stress chips differently.
So Gimlet splits the job and runs each piece on the chip best suited to it. It works with chips from six makers, NVIDIA, AMD, Intel, Arm, Cerebras, and d-Matrix. Its software orchestrates the pieces, and developers see a standard inference API. The company says this delivers either 3 to 10 times higher interactivity or 3 to 10 times more throughput per kilowatt, depending on which target you hold fixed.
It also sells Gimlet Cloud, serverless inference for AI agents, and kforge, which generates optimized GPU kernels from PyTorch code.
The company is moving beyond software into infrastructure. It says it is securing data center capacity and is hiring data center architects.
On September 28, Gimlet announced a partnership with Cerebras to deploy 100 megawatts of Cerebras-powered inference capacity. The first datacenter is expected online later this year, with speeds of up to 3,000 tokens per second. Gimlet will also be a launch partner for Cerebras's next system, the CS-4, in 2027.
On September 4, Gimlet raised a $300 million Series B led by Andreessen Horowitz at a $3 billion valuation. Sapphire Ventures, M12, Arm, Menlo Ventures, Samsung Ventures, and Tiger Global joined, among others. Earlier came a $12 million seed and an $80 million Series A led by Menlo in March, bringing total funding to $392 million. Gimlet says it has tripled its customer base since March and added one of the top three frontier AI labs and one of the top three hyperscalers. It hasn't disclosed revenue or a customer count.
Asgar says the future of inference is vertically integrated and multisilicon, with many chip architectures orchestrated as one system.
10/100·CNeutral+12
m
meme10/25meme
The more personal an AI agent becomes, the more important its boundaries become. 🔐
An assistant may need to know your household preferences. That should not automatically give it access to every private document or every connected service.
AgentKeys is a personal and family agent manager that specializes in permission control. A place to make those boundaries practical, not something you have to explain repeatedly in a chat.
Personal agents are getting better at taking action. Our focus is helping you decide where that action is allowed.
🎛️ Give each agent the access that fits its role. Share selected context and change permissions as your needs change. Those limits belong in the system, not just in a prompt.
An agent can ask for more access. It does not get to approve that request itself.
🧾 Keep permission decisions visible. Review what you have granted, check the record, and revoke access when you no longer want an agent to have it.
For a family, convenience should not require treating everything as shared. For an individual, useful help should not mean handing over everything.
🌱 This is the direction we are building at Heima: personal agents that can do more for you, with boundaries you control.
More help. Your rules.
Follow development:
https://github.com/litentry/agent-terrier
10/100·CNeutral
m
meme10/23meme
If you are typing a prompt straight into a flagship video model and hitting render, you are paying premium prices to find out whether your camera move works.
There's some strategies you can use to save a bit of money, read more on our blog:
https://qtum.ai/blog/dreamzo-previz-workflow
10/100·CNeutral
n
news10/22news
A vision script: “sending an image to a vision model”
vs.
An interoperable Image Analysis Agent: “building an Image Analysis Agent other systems can reliably use.”
A simple script: `image → model → response`
An interoperable Agent adds:
→ `ResourceContent` for attachments → `ChatAcknowledgement` for receipt → Attachment metadata → Agent Chat Protocol → A published protocol manifest
Then comes the analysis logic: `image + prompt → multimodal model → result`
Build it. Get the full breakdown in @Fetch_ai_IL Resource Library ↓ https://innovationlab.fetch.ai/resources/docs/examples/chat-protocol/image-analysis-agent
10/100·CNeutral
m
meme10/22meme
seedance 2.5 vs wan 3.0 is the side-by-side everyone’s posting this week.
same prompt, two clips, two price tags. $3.47 and $1.38 in one test, $1.44 and $0.80 in another. plenty of people say the cheaper one looks more real.
here’s what we think that price gap does to how people make video:
1/ you get more drafts.
at half the cost, a creator who used to render 4 versions now renders 8. the extra 4 are where the good shot usually shows up.
2/ testing moves earlier.
cheap renders turn video into something you sketch with. try the lighting, try the pacing, try the ending, then commit.
3/ comparing models becomes part of the job.
the posts going viral this week are A/B tests. creators are building their own benchmarks: same prompt, same reference, price in the corner.
4/ the brief gets more valuable.
when every model renders a good-looking clip, your prompt, your reference image and your shot list are what make the result recognizably yours. write the identity once, change the shot.
5/ the savings go into length.
wan 3.0 runs up to 30 seconds in one take. cheaper seconds mean more creators will try full scenes, short ads and music video cuts.
our take: falling prices reward the people with the clearest idea of what they want. the cheaper a render gets, the more a good plan is worth.
what’s your go-to video model right now, and what does a clip cost you?
10/100·CNeutral
m
meme10/21meme
New in Venice Studio: Seedance Draft Mode.
Toggle "Draft On" with Seedance selected as model and it will render a fast, economical low-res draft first.
Once you know the prompt works, toggle off Draft mode to generate it in full HD.
Draft Mode saves you credits + time.