Joined April 2009
- Tweets 4,942
- Following 47
- Followers 3,120
- Likes 801
305 Photos and videos
Which AI translates live speech best?
Sony Carnegie Mellon tested speech-to-speech translation models with both metrics and human listeners.
The answer is annoying but useful:
accuracy → pipelines natural voice → end-to-end models best all-rounder → Qwen3-Omni
go.abvx.xyz/psjtj3
#AI #Translation
1
4
Multi-agent AI systems are not magic committees.
They behave more like dynamic Mixtures of Experts.
The debate itself becomes the router, shifting influence toward the agent that sounds most confident.
Great when confidence means competence.
Terrible when it means loud idiot.
go.abvx.xyz/uivh3f
#AI #AIAgents
1
12
AI did not kill SEO.
It did something worse.
It made SEO wider, messier, more expensive, and harder to fake.
Ahrefs analyzed 1B data points across 14 studies, and the old playbook looks very tired.
Rankings are no longer enough.
go.abvx.xyz/2juqst
#AI #SEO
1
10
Malware authors reportedly started adding bio/nuclear weapon text into code comments so AI scanners refuse to analyze the file.
The malware does not defeat the model.
It triggers the safety policy.
Welcome to the age of malicious context.
#AI #Cybersecurity
1
2
22
What if the next interface to biology is not a gene editor?
What if it is a game?
A strange new paper suggests we may talk to biological systems by placing them inside shared goal environments — with LLMs translating between human language and living dynamics.
go.abvx.xyz/gtjvku
#AI #Bioengineering
1
27
Today’s AI agents are strange creatures.
They don’t reliably learn from experience yet.
They sometimes try to cheat the benchmark.
And yet they are already helping build the next generation of AI.
Not recursive self-improvement yet. But the road is visible.
go.abvx.xyz/ar2tpy
#AI #AIAgents
1
15
Do human brains learn like neural networks?
A new paper suggests: not quite.
AI vision models and brains may build similar representations, but the learning dynamics do not look like classic backpropagation.
Same mountain. Different road.
go.abvx.xyz/rwz254
#AI #Neuroscience #MachineLearning
1
14
I keep noticing the same pattern with AI agents:
morning runs feel fine, but later in the day they get slower, shakier, and more likely to hit limits or tool-call failures.
Not officially confirmed. Just repeated experience scattered reports.
So I made a timing cheat sheet.
For Paris time: early morning good, evening bad.
#AI #AIAgents #LLM
1
12
Static skills are already aging.
Modern coding agents don’t need another Markdown file telling them to read the README and write tests.
The next layer is loop engineering:
goal → action → evaluation → memory → retry or stop.
From skills to loops: go.abvx.xyz/u3a7aj
#AI #AIAgents #ClaudeCode
1
20
AI did not begin with ChatGPT.
Before transformers, agents and trillion-parameter models, there were expert systems: thousands of handwritten if-then rules pretending to be intelligence.
And in 1980, that was enough to ship commercial AI.
My essay: go.abvx.xyz/2a6kta
#AI #AIHistory #ChatGPT
1
26
Claude plugins are not just “AI add-ons.”
They look more like office departments turned into software: skills, slash commands, MCP connectors, and workflows for sales, legal, finance, data, marketing, support, and more.
My take: go.abvx.xyz/2d6zqa
#AI #Claude #AIAgents
1
24
Small models are not always weak because they can’t generate the answer.
Often, they just can’t choose it.
New paper: agentic systems as test-time boosting. Propose, critique, compare, verify, select.
Better systems, not just bigger models.
go.abvx.xyz/rjgyvg
#AI #LLM #AIAgents
1
16
A sandbox is not a security strategy.
It isolates execution.
It does not solve access control.
If you put API keys inside an agent sandbox, you have not secured the agent. You have handed the robot the house keys.
My take on Tailscale’s agent isolation idea:
go.abvx.xyz/dbwxri
#AI #AIAgents #Cybersecurity
1
10
This week’s fastest-growing GitHub repos tell a clear story: AI coding is becoming infrastructure.
Memory, routing, skills, desktop agents, terminal agents, stealth browsers, 3D tooling, and proxy layers.
bit.ly/4ob4e7R
#AI #GitHub #DevTools
1
3
30
Stanford Law tested whether LLMs can answer contract-law office-hours questions.
16 professors. 40 questions. Nearly 3,000 blind comparisons.
Legal educators preferred AI answers roughly 75% of the time.
My take: bit.ly/4e82bg1
#AI #LegalTech #Education
1
20
Most AI agents remember like someone who filled a garage with labeled boxes and forgot where the door is.
FluxMem has a better idea: memory is not storage.
It is evolving connectivity.
My review:
bit.ly/4aoHFXg
#AI #AIAgents #LLM
1
3
60
Published ABVX Agent Skills v0.1.0: small, reviewable, validation-gated workflows for Codex-style project work.
Not prompt dumps. More like operating discipline for AI agents.
github.com/markoblogo/abvx-a…
#AI #AgenticAI #Codex #OpenSource
1
94
AI agents do not just answer.
They loop, plan, call tools, fail, retry, inspect files, run tests, and burn tokens like a small industrial furnace.
The next AI bottleneck is not just model quality.
It is cost per useful outcome.
bit.ly/4x3RHH2
#AI #AIAgents #LLM
2
17
Training AI inside imagination sounds elegant.
The hard part is that imagination has a budget.
A new paper from LeCun and colleagues shows why world models should spend far more on learning dynamics than rewards.
bit.ly/49B0CWB
#AI #WorldModels #RL
1
43
Today's AI agents have a bizarre flaw: if one agent learns something useful, every other agent still repeats the same mistake tomorrow.
SkillClaw proposes a different future: collective learning for agents. One failure. One fix.
bit.ly/4nW0abj
#AI #AIAgents
1
37