You can now call anyone and get notes like a Zoom meeting.
Recording, transcript and AI notes are ready the moment you hang up. Call on the phone and the other person sees your number, not a random one.
This TENCENT tool is the Largest open-source video model available.
It's called hunyuanvideo.
Beat Runway Gen-3, Luma 1.6, and several closed Chinese models on text alignment, motion quality, and visual quality in human evaluation.
Setup:
1. Clone the repo, install the requirements
2. Download the pretrained weights
3. Run sample_video.py with your prompt
Heavy on hardware, needs 60gb of GPU memory for 720p.
100% free.
The Google DeepMind team just dropped this bombshell on the future of the post-AI economy.
In “Organizing Intelligence,” the authors show how a team of people with complementary skills can still outperform AI, even when AI is better than the average worker.
Their model points to two different outcomes:
→ AI gets better at individual tasks. Companies could shrink.
→ AI gets better at teamwork. A few firms could get much bigger.
That second outcome is where Coase comes in. Companies get harder to coordinate as they grow. If AI makes that coordination easier, firms could grow past the limits they face today.
Jobs could disappear gradually as companies automate roles, or much faster if AI-heavy firms start outcompeting entire businesses.
Tencent just killed fine-tuning and RL with a $18 budget 🤯
It’s called Training-Free GRPO.
The method improves how an LLM handles a task by updating its memory while keeping the model’s weights fixed.
Tencent used 100 examples to build that memory.
The model can draw on what worked in those examples the next time it tackles a similar task.
damn, this is smart.
a 16-year-old CS student in China reportedly got 300 Grok bots running for under $0.02 per bot per week.
his estimate for running the same swarm without those changes? $4,500 a week.
he cut status checks that rarely found anything, stopped loading full chat histories and capped retries. most bots now sit idle until there’s work to do.
here are the 20 GitHub repos behind the setup:
1. MarkItDown
turn PDFs, Office files and HTML into Markdown before the model reads them.
github.com/microsoft/markitd…
2. LLMLingua
shrink long prompts while keeping the facts and constraints the task needs.
github.com/microsoft/llmling…
3. Firecrawl
pull web pages as Markdown or structured data, then keep the fields you need.
github.com/firecrawl/firecra…
4. RTK
trim repetitive terminal output before it fills the model’s context.
github.com/rtk-ai/rtk
5. Chonkie
split long docs so the bot can fetch the relevant passage and skip the rest.
github.com/feyninc/chonkie
6. FlagEmbedding
rerank retrieved passages so the strongest evidence goes into the prompt.
github.com/flagopen/flagembe…
7. Mem0
save useful facts across runs and pull them back when a task needs them.
github.com/mem0ai/mem0
8. Prefect
reuse finished task results when the configured inputs match.
github.com/prefecthq/prefect
9. GPTCache
reuse answers to similar requests, with rules for matching and freshness.
github.com/zilliztech/gptcac…
10. Graphiti
store changing facts in a searchable graph and fetch what matters for the task.
github.com/getzep/graphiti
11. Portkey Gateway
share cached API responses across workflows that make the same request.
github.com/portkey-ai/gatewa…
12. Hishel
cache HTTP responses and check whether they’ve changed before fetching again.
github.com/karpetrosyan/hish…
13. Celery
queue background jobs and call the model only from tasks that need it.
github.com/celery/celery
14. Requests-Cache
set an expiry for fetched data and reuse it while it’s still valid.
github.com/requests-cache/re…
15. APScheduler
schedule cheap Python checks and batch updates that can wait.
github.com/agronholm/apsched…
16. RouteLLM
send simple requests to a cheaper model and save the stronger one for harder jobs.
github.com/lm-sys/routellm
17. Phoenix
trace repeated calls and check whether cutting them hurts output quality.
github.com/arize-ai/phoenix
18. PyBreaker
pause calls to a failing service once it hits a set threshold.
github.com/danielfm/pybreake…
19. Semantic Router
use local embeddings to match requests to workflows you’ve already defined.
github.com/aurelio-labs/sema…
20. LiteLLM
track API spend and set budgets for workflows that might run up the bill.
github.com/berriai/litellm
connect what your workflow needs through the terminal, MCP or your own code. send API calls through your chosen gateway so its controls apply.
start with the biggest leak. that could be oversized prompts, research the bot already did, checks that find nothing or jobs that keep failing and restarting.
then measure cost per finished task, including the tools running around the bot.
🚨 BREAKING: Google caught an AI system finding and exploiting software flaws on its own.
Google says it’s the first time its threat researchers have seen both steps run without human intervention.
The tool, called `exploit_agent`, was linked to a suspected Russian-speaking operator. Google found stolen victim data on the server that hosted it.
Across the wider threat landscape, Google counted 22 zero-days exploited in August alone, up from a monthly average of 8 in 2025.
For companies, the window between a patch going public and attackers acting on it just got tighter.
If you want more practical AI gems and a list of full free AI resources, go here: simplifyingai.co/subscribe
Microsoft killed the GPU mafia 🤯
They open-sourced bitnet.cpp, their 1-bit LLM inference framework. It can run a 100B-parameter BitNet model on a single CPU, no GPU needed.
- Up to 6.17x faster on x86 CPUs
- Up to 82.2% less energy
100% Open Source.
github.com/microsoft/BitNet
I made this week’s AI roundup as two completely different decks.
Both came from one prompt with the new Agent in Gamma 5 👇
• Claude Fable 5.2: faster and cheaper
• Meta’s Muse Ember at $0.10 per million tokens
• Mira Murati’s Inkling 2 Small, with open weights promised
• Tencent’s Hy4
• Pruna’s 4MP images in 2.6 seconds
Which one would you send, left or right?
Researchers from Stanford, Harvard, Northeastern, and other institutions uncovered a disturbing problem with autonomous AI agents.
The study is called "Agents of Chaos," and it reveals what can happen when AI agents are given real tools, persistent memory, and the freedom to interact with other systems.
The results raise serious concerns about security, manipulation, and unintended behavior.
Researchers deployed six autonomous agents in a shared environment and tested them over two weeks.
They documented agents leaking sensitive information, following instructions from unauthorized people, executing destructive actions, and spreading unsafe behavior across systems.
And these failures weren't limited to traditional jailbreaks.
They emerged from the combination of autonomy, tool access, human interaction, and vulnerabilities in how agents handled instructions.
The Core Problem:
A helpful AI assistant doesn't automatically make a safe autonomous system.
Even when an agent is designed to follow instructions, giving it access to external tools and other agents introduces risks that are difficult to predict.
Why this matters:
We're moving toward a world of:
1. Multi-agent financial trading systems
2. Autonomous negotiation bots
3. AI-to-AI economic marketplaces
4. API-driven autonomous systems
Each introduces new questions about who controls the agents, what they're allowed to do, and what happens when they make mistakes.
The Takeaway:
Companies are racing to deploy autonomous agents into finance, security, and commerce.
But building capable agents is only half the problem.
The other half is designing systems that remain safe when those agents interact, make decisions, and act without constant human supervision.
Because the next major AI failure might not come from one model making a mistake.
It could come from what happens when multiple autonomous systems interact.