Exploring AI. Building in public. Tools, experiments & insights. Same curiosity. Higher possibilities.
Joined July 2026
- Tweets3.6K
- Following171
- Followers223
- Likes6K
12 AI models tested.
0 successful VM escapes.
But here's where AI agent security gets interesting.
Perplexity's SPACE research found that agents could still exploit network-policy weaknesses in its earlier sandbox implementation.
The vulnerabilities were fixed, and follow-up tests found no successful bypasses.
The bigger lesson?
An AI agent sandbox needs more than one security boundary.
You need to test:
1. VM isolation
2. Network egress restrictions
3. Credential access
4. Prompt and task manipulation
"Runs in a sandbox" isn't a security strategy.
It's the beginning of one.
1
2
70
30% vs 16%.
Google DeepMind's AlphaGenome 2 nearly doubled the identification of causal genetic variants in a specific immune-cell benchmark.
Here's why this matters:
Every cell in your body shares the same DNA.
But your brain cells, heart cells and immune cells don't read that DNA in the same way.
AlphaGenome 2 is trained on 107M+ individual cell profiles across 68 tissues.
It can even analyse previously unseen cell states without retraining.
We're moving from predicting what DNA does to understanding WHERE it does it.
AI + biology is getting very interesting. 🧬
1
1
35
Your AI agent doesn't need to generate a paragraph to make a decision.
Google just introduced Gemini Decision 1.
Instead of generating tokens, it can:
→ Evaluate up to 1,000 questions in one pass
→ Process text, images, audio, video and PDFs
→ Return probabilities and confidence scores
→ Deliver decisions in 150–250ms
Pricing: $0.10 per million input tokens. Zero output-token charges.
This could fundamentally change how we build AI agents.
Not every step needs a frontier LLM.
Sometimes, a fast, calibrated decision is all you need.
2
4
151
Riding the hype train
just found out you can screenshot any post directly from the share menu on Android 🤯
Tap the share icon → Screenshot (camera icon) → clean image of just the post.
No more messy screenshots.
How did we all miss this?
2
3
101
Hot take: Evaluating AI products is 100x easier than old SaaS.
Old SaaS = weeks of demos + full team trial just to figure out which one is actually better.
AI today = check the leaderboard, run the benchmarks, look at the evals. Numbers settle it in minutes.
Who else is seeing this shift? Which model are you ranking highest right now and why?
2
3
51
The Untitled Goose Game studio (House House) confirmed it used LLMs “sparingly” near the end of Big Walk’s development mainly to help fix bugs in third-party code.
They estimate well under 0.1% of the game’s code was affected. No generative AI was used for art, writing, or other assets.
A group of IGF 2027 jurors has said they will not vote for any game that used generative AI in any capacity. The disclosure has already triggered review bombs and refund requests on Steam.
This is the new fault line in indie games: is tiny, practical LLM use for debugging acceptable, or does any use disqualify a game?
2
4
171
7 dormant neurons. One hidden trigger. A model that still passes its evals.
Microsoft Research's new “Lazarus” study shows how fewer than 10 edited neurons can turn an AI model into a sleeper agent, without retraining.
In one Nemotron-Terminal-8B experiment, just 7 of 12,000+ neurons were enough to activate a payload when a trigger appeared.
This is a lab-demonstrated technique, not evidence of an active campaign.
But agent security needs to go beyond model evaluations:
→ Verify model provenance
→ Audit model weights
→ Sandbox tool access
→ Require approval for high-impact actions
Passing evals isn't the same as being safe to run with your credentials.
4
7
245
AI coding is moving from one developer + a fleet of agents to an entire team + connected fleets.
Cursor just made Projects multiplayer:
→ Teammates share context, specs, diagrams and decisions
→ Each person gets their own coordinator and worker agents
→ Agents coordinate across teammates to catch overlapping work before it becomes a merge conflict
The real shift isn't “more agents.”
It's shared context + agent-to-agent orchestration.
That changes how teams plan, hand off and ship software.
The next bottleneck isn't agent count.
It's coordination.
1
5
128
Qwen4 just moved from “coming soon” to preview.
Alibaba says Qwen4-Max is now available through Token Plan, with early access on QwenWork, Qoder, Qwen APP and QwenStudio.
The pitch:
• Deeper knowledge
• Richer world understanding
• Stronger coding
• More capable agentic workflows
The launch post hasn't provided public benchmark numbers, so the real test is what it can actually build.
Give it a real repo, a messy front-end task and a multi-step workflow.
A side-by-side demo will matter more than another model announcement.
#Qwen4
3
195
Finally it's here
Starting today, our next frontier model Gemini 4 Carbon is rolling out to Google AI Ultra subscribers on desktop, mobile, and web at gemini.google
It’s our strongest model, built to reason deeply through your most complex work:
✅ Get more thorough, reliable outputs on your toughest tasks
✅ Turn dense data and docs into polished presentations, spreadsheets, and ready-to-use reports with more precise analysis
✅ Create web apps, prototypes, and 3D scenes with stronger spatial understanding for 3D and UI design
2
6
232
MiniMax just entered the image generation arena. 👀
Introducing H3-Image.
16 reference images.
More control over composition and consistency.
From character design to 3D environments and industrial visuals.
The interesting part? It's built to understand what you actually want to create, not just generate pretty pixels.
Image generation is getting serious. 🔥
Can't wait to push this model to its limits.
#H3Image
2
5
196
Grok 4.8 update from Elon Musk today:
> Pre-training is done
> Now in reinforcement learning (evals climbing)
> Entire RL + inference stack rewritten in C/C++
Soft rollout to Grok Bot users later this week
He previously noted it is a 2.5T model and expects it to differ significantly from Grok 4.7.
3
5
191
The open-source X ranking code scores posts as
Score = Σ (weight × predicted probability of each action).
Replies and shares are weighted far higher than likes (reply ≈ 5–20× a like; copy-link share = 20). This creates a feedback loop: content that triggers replies including ragebait and hate gets amplified into trends, while lower-engagement genuine or nuanced posts are suppressed.
A 2026 PNAS study confirmed the algorithm amplifies value-misaligned content largely through reply-driven engagement. X’s then product lead also acknowledged the reply predictor as a major driver of ragebait.
4
4
133
PNAS (Epstein et al.): doi.org/10.1073/pnas.2610388…
2
48
Make your coding agent say who did what before it explains anything.
Zonder Labs hand-labeled Claude Code replies: 78% of sentences were unclear, vs 11% of Stack Overflow answers. Vague wording was the top fault.
Their real example: "The database readiness check sits at the user-session entry boundary."
Add one rule to CLAUDE.md: subject does verb, plain nouns, result first.
7
5
225
