Moez AI retweeted
Follow-up on the 9-byte CLAUDE.md.
Two of you suggested the same guard: fail CI if git stores an agent rules file as a symlink (mode 120000).
It's a one-liner. It caught it on a scratch repo.
What else is in your CI because an agent failed quietly?
Moez AI retweeted
Dubai Future Week 2026 will begin on November 18 under the theme “For the Human of the Future”. The four-day event will take place at Emirates Towers and the Museum of the Future until November 21. It will bring together experts, scientists, business leaders and innovators from around the world to discuss how technology and new ideas will shape the future. The programme will include around 70 talks and panel discussions, 30 workshops, exhibitions, competitions and learning programmes covering areas such as artificial intelligence, robotics, business and education.
The event will also feature projects from 100 student teams, startup presentations and discussions on how new technology can improve everyday life. One of the main highlights will be the AI Grandmaster Challenge, where the winner will receive AED 1 million. Visitors will also be able to explore interactive activities involving technology, art, food and culture. Organised by Dubai Future Foundation under the patronage of Sheikh Hamdan bin Mohammed bin Rashid Al Maktoum, the event aims to bring people together to share ideas and explore future opportunities. Photo: WAM
#DubaiFutureWeek #Dubai #UAE #ArtificialIntelligence #Innovation #Technology #ArabTimesNews
what in the good lord is this!!! 'speed of light'?!! I test endpoints on my own pipelines, not on launch hype. If it's public I'll benchmark it.
What if you could let your computer read and think at the speed of light?
Sage2 pushes past the decision model. We are now combining extraction, streaming, and decisions through one endpoint!!!
whats your favorite example from the video?
10
I agree with the delay logic. Those departures since 2024 were a loud signal. Embedded assessments only carry weight if the findings get published in full. Without transparency, there is no absolute safety.
I think the reason OpenAI has delayed gpt-6.2 for so long, is to finally pay off years of accumulated security debt.
no other frontier lab had so many researchers leave for safety concerns as far back as 2024.
and after the cyber incidents, we learned from evaluators that OpenAI had faulty monitoring (METR, Redwood) and sandboxing (Reuters).
then Reuters reported that Tibo reassigned 25% of OpenAI’s product engineers to work on its security infrastructure. (ironically, many of the researchers who left OAI said that it was prioritizing products over safety.)
the last thing OpenAI wants is another wave of cyber incidents and they’re finally taking this seriously.
5
Preserving the original result and revision history while doctors edit contours is the part I care about most. Clinical approval depends on full audit trails, and most systems throw the original output away. This is built for real validation.
🇨🇳BAAI has released Cardiac Agent v1.5, an AI agent for cardiac MRI analysis.
It can analyze 10 types of MRI sequences, quantify heart function and myocardial abnormalities, and provide diagnostic support across 17 categories.
It also checks findings against 18 selected criteria from 4 clinical guidelines and expert consensus documents, flagging inconsistencies for doctors to review.
In a 100-case clinical evaluation:
• Primary diagnosis agreement: 92%, vs. 64% for GPT-6 Astra.
• LVEF mean absolute error: 5.7 percentage points, vs. 9.7 for GPT-6 Astra.
Doctors can edit segmentation contours, with the system recalculating measurements while preserving the original results and revision history.
3
👀 DeepSeek just outperformed Claude Opus 5.5 in an AI agent coordination benchmark.
We tested a different architecture with our friends at @agent__swarm, adding explicit rules and Clef, a decision model from @Cloudflare, to support agent coordination.
The results?
DeepSeek V4.1 Flash went from 83% to 100% success on evaluated runs. Claude Opus 5.5 scored 93%.
The sample is still small, and the runs weren't identical, so there's more testing to do. But the result is promising: we may be able to improve agent performance by changing the architecture, without always needing a larger model.
Huge credit to @voidisomorphism and @ecura for running the experiment.
Next, we're expanding the tests to more software engineering tasks.
We'll publish a technical article and video breaking down the architecture, methodology and results next days.
Full benchmark 👇
api.desplega.agent-swarm.dev…
Inkling 2 Small is a great model indeed. I would add K2 Horizon 375B to your evals too, fully open weights, code and training data, 47 on Artificial Analysis. Try it!
Really bullish on Thinking Machines (and some others) for an American open source comeback. Inkling 2 Small is a great model, and also it feels unique in how it thinks. I think you should try it. Let's go!
6
Fable at $49.68 and 87 minutes for 11 steps, I am not impressed! Opus ran the same task at $30 in 1h02. I want to know how many shots until the swarm fails to reassemble.
Fable 5.2 vs Opus 5.5 Humanoid Worm Swarm (Unreal Engine + Blender)
The prompt was to build a humanoid creature made of hundreds of living worms. When you shoot it, the worms fly off, crawl back, and rebuild the body.
API Cost:
Fable 5.2 (High) = $49.68
Opus 5.5 (High) = $30.08
Duration:
Fable 5.2 (High) = 1h 27m
Opus 5.5 (High) = 1h 2m
1
15
Moez AI retweeted
Archify converts natural language descriptions into interactive architecture diagrams.
github.com/tt-a1i/archify
Moez AI retweeted
🔗 GitHub: github.com/goat-si/goat
---
✉️ If you’re into AI, ML, agents, and building real systems, join my newsletter (it’s free): dankornas.substack.com
No signs yet. Optimizer tweaks that look real at 1B regress at 70B in my own runs. The verifiable-proxy argument assumes a cheap feedback loop nobody has built for training algorithms.
Replying to @dwarkesh_sp
I am yet to see it come up with a better optimization algorithm for say a frontier model, do you see signs where it can beat current SOTA, prime intellect results showed it was very much stuck in a local minima.
2
196
I'd weight operator jurisdiction above data residency in the scoring. Abu Dhabi is a live case: sovereign compute where the infrastructure, the operator, and the legal jurisdiction sit in one place. None of the seven can offer that.
🔎 SOVEREIGN OR NOT, HERE I COME! ☁️
Every hyperscaler is king of its own domain. 👑
But how sovereign are they when you put them all on the same scoreboard?
We scored and ranked:
@awscloud Web Services (AWS)
@Azure Azure
@googlecloud Cloud
@Oracle (OCI)
@IBMcloud Cloud
@alibaba_cloud Cloud
@Huawei Cloud
7 hyperscalers. 40 offerings. 5 buyer regions. One very uncomfortable buyer's guide. 💣
And the findings?
🔌 Nobody has really tested the off switch.
⚖️ The CLOUD Act doesn't care where your data center sits.
💸 Your exit strategy could cost you up to $79K.
🇪🇺 Europe's sovereignty framework still leaves gaps.
The plot twist? The most sovereign cloud is the one the hyperscaler itself can't reach.
This isn't a purity test. It's about knowing what control you're giving up, what risk you're taking, and what it'll cost you to walk away.
Don't like your grade? Publish the document. We'll re-score it. 😉
📖 Sovereign or Not? The 2026 Hyperscaler Buyer's Guide lnkd.in/e-dTUwuf
🎙️ David Vellante and I unpack the findings in the latest @Breaking Analysis from SiliconANGLE & theCUBE.
hashtag#SovereignAI hashtag#CloudSovereignty hashtag#AI hashtag#Hyperscalers hashtag#BreakingAnalysis John Furrier
lnkd.in/p/eKK5Scfd
1
55
Every pause proposal I have reviewed lacks an enforcement mechanism. Training compute, weights and researchers are mobile across jurisdictions, so the buildout simply relocates. Governance only means something with global adoption.
Bernie Sanders wants to pause AI development, which is honestly stupid. You can’t pause AI. Who else is going to take the crown?
14
Agree on the core point. I publish fine-tuned open models regularly and the bottleneck moved long ago from capability to judgment: which base model, which dataset, which use case. Your sourcing engines follow the same pattern.
When anything can be built, the real work is figuring out what to build.
@StanfordGSB asked a group of alumni how AI is changing our work and how to stay relevant as it does.
Two observations:
First, today almost anyone can build almost anything. Far fewer people know what’s actually worth building.
I think it takes two things together. One is building an encyclopedic knowledge of a space, the way you do only after exploring rabbit hole after rabbit hole. The other is tinkering at the frontier with these tools, exploring the boundaries of what’s possible.
Second, you can never stop learning and adapting the way you work. For example, at Foundation Capital, we’ve built always-on AI sourcing engines to help us find great founders building in niche Slack and Discord communities. I expect these tools to continue improving as models and infrastructure do.
Glad to be included alongside @thefriley, Kate Jensen, @ravi_lsvp, @federicoantoni and @jakesaper.
More here: lnkd.in/gvfBqEaH
20
Moez AI retweeted
Fine-tuning an LLM shouldn’t mean stitching together a new training pipeline.
Axolotl is an LLM fine-tuning framework for builders adapting models to their own data and tasks.
It helps you keep experiments in one workflow by reusing a single YAML configuration across dataset preprocessing, training, evaluation, quantization, and inference.
Key features:
• Multiple training methods – choose full fine-tuning, LoRA, or QLoRA for your experiment
• Flexible data loading – use datasets from local files, Hugging Face, or cloud storage
• Preference tuning – supports DPO, IPO, KTO, and ORPO for training on preference data
• Multi-GPU and multi-node training – includes FSDP2 and DeepSpeed support, plus Torchrun and Ray options
• Example configs included – fetch the examples and start from a provided LoRA training configuration instead of a blank file
It’s open-source (Apache 2.0 license).
Link in the reply 👇
Moez AI retweeted
Formal verification is a hot topic!
I wasn't expecting so much interest in my seL4 thread.
I'm glad to see serious discussion about what FV can guarantee, its limits, and why adversarial research still matters.
Welcome new followers!
More interesting security research ahead.
Power access is the constraint I watch closest. Transformer lead times above 160 weeks force fresh capital raises. Sites with dedicated baseload, like Barakah's 5.6 GW in Abu Dhabi, remove that dependency.
Nscale warns it needs fresh financing as it prepares for a New York flotation, amid growing demand for AI infrastructure. agbi.com/ai/2026/11/abu-dhab…
16






















































