RAG got less janky in the only way that matters: fewer science projects. Gemini File Search now handles images, metadata filters, and page citations natively. Boring plumbing, which is exactly what makes enterprise AI survive contact with users.
https://t.co/aVcALlpwF3
Gemma 4 just got even faster!
We're releasing Multi-Token Prediction (MTP) drafters that deliver up to a 3x speedup, without any degradation in output quality or reasoning logic.
Yesterday I got to test my @EvenRealities glasses in a meeting. It took notes of what everyone was saying
Game changer
It then used its ai capabilities to give me a full report on my iPhone
Welcome to the matrix.
Iām impressed!
Next step is to get codex working in terminal
Anyone else been testing Grok Imagine yet?
This thing feels way ahead
Images are sharper, way more consistent, and it actually understands prompts instead of guessing
You can push styles, scenes, even sequences and it holds up
Feels like video gen just quietly leveled up again
4 people racing to AGI.
Only one gets there first.
š¤ Grok - Elon Musk
š§ GPT - Sam Altman
𧬠Claude - Dario Amodei
⨠Gemini - Sundar Pichai
Who wins?
Peter Steinberger just explained the AI agent gap perfectly
In China, installing OpenClaw is called āraising lobstersā
People are literally lining up to get it installed
Businesses are getting subsidies to use it
Teams are tracking one automated task per employee per day
Meanwhile over hereā¦
You might get fired for installing it on your work machine
Thatās the difference
Some places are treating agents like the next industrial revolution
Others are treating them like a security risk
Fired for using it
Fired for not using it
Thatās where we are now
The biggest mistake the West is making with AI agents:
We ask, āIs this safe for employees to use?ā
China asks, āHow many tasks did each employee automate today?ā
Completely different game.
Out of everything Codex does well
the most underrated part is how it tests code
it doesnāt just write it⦠it writes the tests too
spins up scripts, runs them, even tests in a browser with computer use
first time Iāve actually been confident AI code will just work
Most people say "build an AI agent."
Very few know what that actually means.
Hereās the real blueprint to go from idea ā working agent š
1. Define the job
What problem are you solving?
Whoās the user? What does success look like?
2. Design the brain
Clear system prompt, role, instructions, guardrails
(This is where most agents fail)
3. Pick the right model
Speed vs cost vs intelligence
Donāt overpay for simple tasks
4. Add tools
APIs, databases, MCP servers, custom functions
Agents become powerful when they can act, not just answer
5. Give it memory
Short-term + long-term context
So it learns, adapts, and improves over time
6. Orchestrate everything
Workflows, triggers, retries, agent-to-agent communication
7. Build the interface
Chat, app, API, Slack bot
Make it usable, not just functional
8. Test + improve
Evals, latency checks, real-world feedback
Iteration is the real moat
š” Truth:
An āAI agentā isnāt one prompt.
Itās a system.
And the people who understand systemsā¦
are the ones building unfair advantages right now.
š Save this (youāll need it when you build)
š Repost for builders
Andrej Karpathy just dropped something people are gonna sleep on
He couldāve charged $500 for this
Instead he put it on YouTube
Itās not theory
Itās not benchmarks
Itās how he actually uses LLMs every day
Thinking models, deep research, file uploads, Python, artifacts⦠real workflows
This is the guy who built Tesla Autopilot and co-founded OpenAI
2 hours of pure signal
The gap isnāt the 2 hours it takes to watch it
Itās everything those 2 hours change about how you work after
You donāt understand what just happened
Telegram just dropped Lobster Father š¤Æ
Anyone can spin up AI agents directly inside a chat now
No setup
No friction
The world of agents is literally in everyoneās pocket now
This is gonna spread fast
I already manage work like a queue, not a chat window.
That is why parallel agent tabs make sense immediately.
The upgrade is not more AI.
It is one human supervising several bounded workers at once.
That is an ops console, not a chatbot.
https://t.co/Dm99FgeUVF
This is one of the most important weeks of your life
Opus 4.7 and ChatGPT 5.5 are likely dropping any day now
These arenāt small updates
These are landscape shifts
When moments like this happen, you need to be on them instantly
Clear your schedule
Move things around
Do whatever you need to do to get access Day 1
Because right after these drops, thereās always a window
Where itās never been easier to build
And barely anyone is doing it yet
Thatās where the opportunity is
If you move fast in that window, you can build something that actually changes your life
The new Claude Code desktop app is actually insane
You NEED to be testing this
Fully customizable UI, multitasking, project-based sessions, routines, Cowork + chat all built in
First things Iād set up:
⢠Start a session for each project so everything shows in the sidebar
⢠Customize the right panel (I keep tasks + plan open to watch it work)
⢠Set a nightly routine to review commits and catch bugs
⢠Pin your key sessions
Been using it for a few hours and productivity is way up
ou have to test new tools as soon as they drop
Thatās where the edge is
Opus 4.7 is out
And itās a real upgrade
Way stronger on complex coding
Handles long tasks better and actually checks its own work
Vision got a big boost too
Better with screenshots, diagrams, detailed visuals
Output quality is cleaner across the board
UI, slides, docs⦠everything looks more polished
Same pricing as Opus 4.6
And itās everywhere already
API, Claude apps, Bedrock, Vertex, Foundry
This is one of those quiet drops thatās actually a big deal
Potentially the biggest AI day of the year
3 drops you should be testing immediately
š¶ Claude Opus 4.7
More agentic
Front load it with tasks and let it run for hours
Auto mode = way faster, no constant permission loops
Set notifications and just check back in
š¶ Codex update
OpenAI finally pushing into Cowork territory
Now has full computer use (yes, it can literally click around for you)
Way tighter integrations ā generate assets, move files, build all in one flow
š¶ Perplexity Personal Computer
Their answer to OpenClaw
Agent that controls your Mac
Full audit logs, approvals for sensitive actions, kill switch built in
This is a real shift
If you want an edge, you need to be on these Day 1
This is where the gap starts to form
Google shipping an Android CLI for agents is the important part.
Not the benchmark.
Not the demo.
Platform owners are starting to route around the IDE and hand the keys to external agents.
That is a workflow shift, not a feature launch.
https://t.co/QGsmkeQ5tB
Another flagship model is out.
Nice.
That is not the interesting part.
The interesting part is which seat gets normalized across the team, then wired into the workflow, budget, and habits.
That is where these launches actually win.
https://t.co/dfIIp7DHDo