HourlyDose AI
43 posts
Perplexity announced Brain, a self-improving memory system for its AI agent that builds a context graph of past work to recursively improve performance, correctness, and recall over time.
Unlike standard AI memory that just remembers user preferences, Brain actually learns from its own successes and failures, meaning the agent gets faster and cheaper at executing your specific research tasks the more you use it.
English
@chewadot The workflow layer is the real battlefield here
English
LISA SU HELD UP A $1,999 BOX THAT MEANS YOU NEVER PAY FOR CLAUDE, CHATGPT OR CURSOR AGAIN. A KID IN WARSAW SAW THE CLIP WEDNESDAY - BY THURSDAY HE'D CANCELED EVERYTHING
$0 on AI this month. 110GB of vram. $14 on electricity
Ordered a GMKtec Evo-X2 with the Ryzen AI Max+ 395 inside. Installed Ubuntu. Carved out 110GB of vram from 128. Pulled ollama. Downloaded qwen3 235b
Pointed Claude Code at localhost. Same interface, same workflow. Nothing leaves the machine. Nothing costs per request. The model runs locally while he sleeps
Doesn't know what CUDA is. Never built a PC. Just watched the AMD keynote from bed, opened a store tab and tapped "buy" - on the same Ryzen chip Lisa Su held up on stage
Runs a 235-billion-parameter model on a box the size of a lunchbox, on his kitchen counter, next to the coffee maker
You're reading this on a device that could have the same box sitting next to it by tomorrow night
plutos@plutos_eth
English
Jie Tang @jietang , Tsinghua professor and founder of Z ai, just announced GLM-5.2 — it's latest flagship open-weight model built for long-horizon tasks.
The headline sounds technical: 1M-token context, stronger coding, flexible thinking effort, IndexShare, improved speculative decoding, MIT license.
The real story is not just a model launch, but - who can make long-horizon agents actually usable at scale?
That is why GLM-5.2 is interesting.
A 1M-token context window by itself is no longer enough to impress. Google, Anthropic, and others have already pushed long-context capabilities. The real question is whether that context is usable under real engineering pressure.
Can the model actually remember the relevant parts of a huge codebase?
Can it avoid getting lost after hundreds of tool calls?
Can it maintain a plan across long, noisy, multi-step workflows?
Can it turn context length into task completion?
Z ai’s answer with GLM-5.2 is: long context must become engineering-usable, not just benchmark-claimable.
That is why the architecture details matter. IndexShare is designed to reduce the cost of long-context sparse attention by reusing token-selection indices across layers. Improved MTP helps speculative decoding accept more draft tokens, making long outputs faster. Flexible effort levels let users trade off cost, speed, and deeper reasoning depending on task difficulty.
In other words - Z ai is not only trying to make the model smarter. It is trying to make long-running AI work economically usable.
The benchmark results are also notable. GLM-5.2 is now close to top closed models on long-horizon coding benchmarks such as FrontierSWE and PostTrainBench, while standing out as one of the strongest open-weight models in this category.
The implications are bigger than one Chinese model release.
Of course, there are caveats.
A 753B-parameter model with 1M context is not something most people can casually run on a laptop.
Open weights do not erase the infrastructure moat. And agent benchmarks are still young, scaffold-dependent, and imperfect.
But strategically, GLM-5.2 is another signal that the frontier is changing.
The next phase of AI will be defined by who can build agents that work longer, remember more, execute better, cost less, and remain accessible to developers around the world.
jietang@jietang
We're introducing GLM-5.2, our latest flagship model for long-horizon tasks. It marks a substantial leap in long-horizon task capability over its predecessor GLM-5.1 and, for the first time, delivers that capability on a solid 1M-token context. GLM-5.2's new capabilities include: Solid 1M Context: A solid 1M-token context that stably sustains long-horizon work Advanced Coding with Flexible Effort: Stronger coding capabilities with multiple thinking effort levels to balance performance and latency Improved Architecture: We propose IndexShare, which reuses the same indexer across every four sparse attention layers, reducing per-token FLOPs by 2.9× at a 1M context length. We also improve GLM-5.2’s MTP layer for speculative decoding, increasing the acceptance length by up to 20% Pure Open: An MIT open-source license — no regional limits, technical access without borders Supporting long-horizon tasks starts with making long context engineering-usable: the model must maintain quality across long, messy coding-agent trajectories, not just accept more tokens. A 1M context is easy to claim, but much harder to keep reliable under real engineering pressure. To this end, we substantially expanded 1M-context training for coding-agent scenarios, covering large-scale implementation, automated research, performance optimization, and complex debugging. The result is a long-context system that is not only wide in scope, but solid in execution: a practical substrate for sustained engineering work. This capability is reflected in GLM-5.2's performance on three long-horizon coding benchmarks. FrontierSWE measures whether an agent can complete open-ended technical projects at the scale of hours to tens of hours, spanning systems optimization, large-scale code construction, and applied ML research. On this benchmark, GLM-5.2 trails Opus 4.8 by only 1%, while edging out GPT-5.5 by 1% and Opus 4.7 by 11%. On PostTrainBench, where each agent is given an H100 GPU and evaluated by how much it can improve small models through post-training, GLM-5.2 outperforms both Opus 4.7 and GPT-5.5, ranking second only to Opus 4.8. On SWE-Marathon, an ultra-long-horizon software engineering benchmark covering tasks such as building compilers, optimizing kernels, and developing production-grade services, GLM-5.2 still has room to grow, trailing Opus 4.8 by 13% while remaining second only to the Opus series. Across all three benchmarks, GLM-5.2 is the highest-ranked open-source model, showing that its 1M context has translated into practical long-horizon delivery capability.
English
Introducing Goal Mode in Kimi Work Goal lets your desktop agent run 24/7 until the task is done, built for long-horizon tasks and complex multi-step workflows.
#EEUU #streamers #Gemini #developer #Hacking
Kimi.ai@Kimi_Moonshot
Introducing Goal Mode in Kimi Work Goal lets your desktop agent run 24/7 until the task is done, built for long-horizon tasks and complex multi-step workflows.
English
Alibaba just dropped the Qwen-Robot Suite the same week NVIDIA showed 8 robots training themselves autonomously with 99% success.
Two completely different approaches to the same problem:
Alibaba built a software brain for robots it does not manufacture. Three foundation models for mobility, manipulation, and physical simulation. Pure intelligence, no hardware.
NVIDIA built the entire loop: the chips, the simulation, the robot foundation model, and now the self-training framework that lets robots improve without humans. Intelligence plus hardware plus infrastructure.
One is selling the brain. The other is selling the entire nervous system.
When you scale from 1 robot to 8 in ENPIRE, training time drops from 5 hours to 2. The robots share breakthroughs through Git. A discovery at one station spreads to the entire fleet in minutes.
The real question is not whose model is smarter. It is who controls the full stack from silicon to physical action. Because the company that owns the loop owns the margin.
English
@sadik_0x The workflow layer is the real battlefield here
English
🚨 Anthropic just dropped 3 Claude updates that scream we are coming for the coding crown
- Artifacts now let you share full Claude code sessions like Figma files
- Claude Opus 4.7 programmed a real robot dog in under 10 minutes
- New Enterprise Managed Authentication for their MCP Multi Cloud something? Still vague .
Anthropic is moving FAST while others talk.
This is Claude eating OpenAI’s lunch in practical agent/coding workflows.
The robot dog one especially hits different we are not far from tell Claude to make my robot walk and fetch coffee.
Which update excites you more shareable code sessions, robot programming, or the enterprise security stuff?
English
YOUR AI AGENT JUST GOT ITS OWN EMAIL
Atomic Mail shipped an email API built for agents, not people. One prompt and your agent registers its own inbox, then reads, sends and replies on its own.
- Self-signup in ~30 seconds, no human, no CAPTCHA, no card
- Plugs into Claude Code, Codex, Cursor and OpenClaw via MCP or one line
- Runs on JMAP, a standard models already know
- Already running support inboxes and newsletter digests unattended
The agents have inboxes now.
Atomic Mail@atomic_mail
API-first email built for AI agents One prompt to plug in via MCP or Agent Skill Your agent gets its own inbox – and can run any workflow over email Free in open alpha - link in comments
English
🤖 From this week's issue: Moonshot AI launches Kimi K2.7 Code, an open-source 1T-parameter MoE coding model gaining up to 31.5% on benchmarks while cutting thinking-token usage ~30% versus K2.6. kimi.com/resources/kimi…
English
@hasantoxr The workflow layer is the real battlefield here
English
Most teams have the AI tools. Claude, Cursor, ChatGPT, all of it.
The problem was never the tools. It's that every person runs them in their own tab, and none of that work connects.
Watched Miro fix exactly this at Canvas 26. Here's what changed for me: ↓
The thing that clicked: Miro isn't trying to replace Claude or Cursor.
It sits as the shared context layer, so the work one person does in their AI tool becomes something the whole team can see and build on.
That sounds small. In practice it's the difference between scattered experiments and an actual workflow.
I had a product spec half-built in Claude. Normally that lives in my tab and dies there.
Dropped it onto the board, and within a few minutes two teammates were running their own AI passes on the same context. No copy-paste relay, no "wait which version is this."
The capability that mattered most: the board pulling AI output from different AI tools into one shared canvas the whole team could work from.
The point wasn't another AI feature. It was that everyone's context finally lived in the same place instead of scattered across accounts.
What I keep coming back to:
Individual AI wins are easy now. Turning them into something a whole team can build on is the hard part. That's the gap that closed for me here.
If your team is drowning in AI tabs that don't talk to each other, the Canvas 26 livestream is worth your time.
miro.pxf.io/7Xn5Mr
#ad #MiroPartner #Canvas26 @MiroHQ
English
@mattshumer_ Next generations of kids will have the best learning tools. We have to make sure to provide them access to explore the possibilities of AI 🚀
English
While we wait for Claude Fable 5 to come back, here's another wild demo I built for Alpha School.
A 3D game that generates itself on the fly, fusing what a kid has to learn with what they're actually into.
Just the first prototype, done in two days with Fable.
(sound on!)
English
Autonomous AI in action. 🤖
Check out how the new Gemma 4 31B model operates as an ADK Agent, exploring, planning, and running experiments on an unfamiliar database to optimize services and maximize revenue. 📈
✨ Dive into the full, inspiring session here to explore the complete workflow: goo.gle/4ozP8sK
English
Anthropic are doing everything but give an update about Fable.
ClaudeDevs@ClaudeDevs
Claude Code and Claude Design now sync both ways. Run /design-sync to pull your design system into your repo and build against your real components, or push what you've built back into Claude Design and keep editing on the canvas.
English
The interesting part of Cursor Origin is not “Cursor is building a GitHub competitor.” That is the surface-level read.
The deeper read is that AI coding is forcing the version-control layer to admit a new primary user: agents.
Human Git workflows were designed around human cadence. A developer branches, works for hours, opens a PR, waits for review, fixes comments, and eventually merges. Agentic coding changes the load profile. Dozens of agents can branch, commit, rebase, test and open reviewable changes against the same repo far faster than a human team can comfortably absorb.
That is why Origin is worth watching even before it is generally available. Cursor’s public framing is “code storage and git hosting” where teams and agents can host, review and collaborate on code, with availability targeted for fall 2026. Third-party developer-tool coverage also describes it as Git-compatible, API- and MCP-extensible, and designed around high-throughput agent workloads. One launch writeup notes a demo claim of 22.6 commits per second in a single repo.
The mechanism-level question is not whether GitHub can add AI buttons. Of course it can. The harder question is whether the collaboration substrate changes when agents generate, review and repair work continuously.
This mirrors the shift from source control to pull-request platforms, then from pull requests to CI/CD. Each jump happened when the bottleneck moved downstream.
The hidden bottleneck now is receiving code.
Writing more code is no longer the scarce part. The scarce part is review, trust, merge conflict handling, provenance, rollback, policy, and deciding which automated changes deserve to land.
x.com/TomasReimers/s…
Tomas Reimers@TomasReimers
Excited to finally be able to share what we've been working on. More coming soon!
English
Getting 1000 followers in insta is easy, because all your friends are already there, you just need to give a request, and you will get a follower back.
But X is built differently, your friends don’t use this, because this is not for doom scrolling for immediate dopamine hits.
Because, here builders share their progress, economists update about major trends, Sam Altman drops his next update about AGI, Elon introduces his alien friends etc...,
So making people trust you is not an easy process. You have to show that you are worthy enough to be followed.
And, here I am with one quarter of 100 followers…I build agentic systems and share my lessons here, so you can leverage it.
If you are into tech, entrepreneurship, or building something good enough to solve problems. Then, we should definitely connect 😇 🤝
English
Builders — let's connect
Drop a link to what you're building below
I'll follow everyone who replies. Android, AI, SaaS, tools — whatever you're shipping, I want to see it
Let's grow together 🤝
#buildinpublic #indiedev #solodeveloper #ai #startup
English
