Building the AI Workspace for Deep Work @headgymai and Enterprise Swarm at StreamZero. Mostly post about Applied AI, Enterprise Arch., and UX.
Dubai / Las Palmas
Joined March 2009
- Tweets1.5K
- Following1.2K
- Followers524
- Likes1.4K
Listen to Enterprise Intelligence | Moving AI from Experimentation to Execution on Spotify for Creators open.spotify.com/episode/3kd…
Over the past 4 weeks I have not used a single top-tier model for coding(and I write a lot of code).
The models I have been using for coding over the past 4 weeks(and mostly for free) have been Minimax M2.5, BigPickle(fine tune by OpenCode of GLM-4.6 or Minimax2.5) and GLM 5.1.
Over 1000s of tasks I used GPT-5.3 Codex and Opus 4.7 only once. And that one time neither of them could solve the task - I fixed the issue 'manually' in <15 minutes.
The task was simple enough - cache, warmup & reuse model weights for Docling.
Frankly speaking with current harnesses, documentation MCP servers, fine tuning and reasoning top tier models are often overkill.
This is why I am in the UAE. To move things fast, you need to show up.
Under the directives of the President of the UAE, we launch a new government model. Within two years, 50% of government sectors, services, and operations will run on Agentic AI, making the UAE the first government globally to operate at this scale through autonomous systems.
AI is no longer a tool. It analyses, decides, executes, and improves in real time. It will become our executive partner to enhance services, accelerate decisions, and raise efficiency.
This transformation has a clear timeline. Two years. Performance across government will be measured by speed of adoption, quality of implementation, and mastery of AI in redesigning government work.
We are investing in our people. Every federal employee will be trained to master AI, building one of the world’s strongest capabilities in AI-driven government.
Implementation will be overseen by Sheikh Mansour bin Zayed, with a dedicated taskforce chaired by Mohammad Al Gergawi driving execution.
The world is changing. Technology is accelerating. Our principle remains constant. People come first. Our goal is a government that is faster, more responsive, and more impactful.
"OpenClaw is probably the single most important software release ever,” said Jensen Huang. So far, public interest suggests otherwise.
Search volume for OpenClaw is now approaching that of @n8n_io .
That said, Huang may not be wrong about its long-term trajectory. What OpenClaw illustrates, above all, is how weak feature engineering and security issues can hold back even genuinely important innovation.
In a recent survey of tech founders, the takeaway was blunt: “We don’t have time to spend two weeks setting something up, only to then deal with security holes.”
I compare OpenClaw to tools like n8n because that is where it seems to fit best: enabling digital agencies and solopreneurs to build powerful automations for SMEs in sandboxed environments. It is not quite the DIY “lifehacker” use case many people imagine.
DESIGN.md is a markdown file which encapsulates the core design principles of a website. Coupled with an agent(in this case HeadGym's General Agent) it spits out designs/HTML which follow the guides of the original. The colour palette, font styles, sizes, and borders etc.
What is surprising, and revealing is how well it works even with a general agent that is not tuned for such work.
Coupled with a design system I can imagine this can be truly powerful(not difficult to implement).
In this example the HeadGym agent reads the rules from the DESIGN.md and was provided a screenshot of a StreamZero Admin screen and asked to build a replica. The DESIGN.md is based on The Verge's design.
This site has a large collection of these files: getdesign.md/
PS: Would never do an admin dashboard which is so loud.
Tell me you are running out of ideas, without telling me you running out of ideas.
Perplexity is giving away $1 Million in seed investment to whoever can build a billion-dollar company in 8 weeks using Perplexity Computer as your primary tool.
With $1M seed I would happily build a better version of Perplexity computer in 8 weeks . Then multiply it by 26 and call it ARR. Voila, another uniCON is born. 🤣
A must watch!
New video: Y Combinator lets you cross the line.
The adults in the room aren't going to stop you from doing something seriously wrong. Young founders need to be aware of that.
youtube.com/watch?v=ptT_LGfT…
People downloading skills that can trade crypto. What can go wrong?
Minara Skill commands (in case you need it)
Account & Wallet
- login / logout / account
- balance — USDC/USDT balance across spot and perps
- assets — view wallet assets
- deposit / withdraw / transfer
Trading
- swap — cross-chain spot trading
- limit-order (lo) — limit orders (create / list / cancel)
- perps order — place a perps order
- perps close — close a position
- perps cancel — cancel orders
- perps leverage — update leverage
Perps Account
- perps wallets — sub-wallets with balances and status
- perps positions — open positions
- perps trades — fill history
- perps deposit / withdraw / transfer / sweep
- perps create-wallet / rename-wallet
Autopilot
- perps autopilot (ap) — manage AI autopilot trading strategy
AI & Analysis
- chat / ask — quick AI chat
- research — deep research mode
- perps ask — get AI analysis for long/short direction
Market Discovery
- discover trending — trending tokens or stocks
- discover search — search tokens or stocks
- discover fear-greed — Fear & Greed Index
- discover btc-metrics — Bitcoin market metrics
github.com/Minara-AI/skills
Looks like Anthropic’s marketing team is in overdrive, hyping Mythos with some pretty tall claims - yes, the new model that’s supposedly so powerful they won’t release it.
A new analysis found that small, cheap, open-weight models recovered much of the same reasoning behind Mythos’s showcase vulnerability exploits. Not just one model, 8/8 models detected the flagship FreeBSD bug!
On some security reasoning tasks, small open models beat larger frontier systems.
The implication is hard to miss: the moat is the ‘wrapper’, not the model.
This increasingly feels like marketing for their Project Glasswing. Strategy: if your wrapper is open to scrutiny, the safest play is to suggest the underlying model has capabilities no one can verify.
Read the full article by Stanislav Fort of AISLE™ : aisle.com/blog/ai-cybersecur…
We are deluding ourselves if we think deploying software that runs unsupervised, dynamically generated code on always-on personal servers exposed to the internet is anywhere near mainstream.
And if we thought either of those issues was significant, just wait for the skills apocalypse: a natural-language virus with a built-in distribution network, vulnerable user base, and indeterminate auditability.
OpenClaw, Hermes, and similar tools are becoming the new N8N, enabling ‘agencies’ to play the agentic game by hand-holding users through setup and extension, before eventually moving them to the cloud.
That opens up serious liability questions as well.
To paraphrase my favourite author, William Burroughs - Language is the virus.
Not good enough, misses the obvious. Am not going to write a paper but the better is.
Singleton code base hosting, multi-user/multi-agent async interface to sequetial instruction queue, intelligent prioritisation(optional) at hub. Exponential reduction of concurrency chaos.
Mature hubs will maintain temporal change and code graph. Is a DAG with all it's power of replay / fork etc.
NEW research from CMU.
(bookmark this one)
The biggest unlock in coding agents is understanding strategies for how to run them asynchronously.
Simply giving a single agent more iterations helps, but does not scale well.
And multi-agent research shows that coordination > compute.
A new paper from CMU proves this with a practical multi-agent system.
CAID (Centralized Asynchronous Isolated Delegation) borrows proven human SWE practices: a manager builds a dependency graph, delegates tasks to engineer agents who work in isolated git worktrees, execute concurrently, self-verify with tests, and integrate via git merge.
CAID improves accuracy over single-agent baselines by 26.7% absolute on paper reproduction tasks (PaperBench) and 14.3% on the Python library
development tasks (Commit0).
The key insight is that isolation plus explicit integration beats both single-agent scaling and naive multi-agent approaches.
For long-horizon software engineering tasks, multi-agent coordination using git-native primitives should be the default strategy, not a fallback.
Paper: arxiv.org/abs/2603.21489
Learn to build effective AI agents in our academy: academy.dair.ai/
Seeing many MetaHarness+AutoResearch babies lately. Soon we will have an AutoResearch builder for everything 😂 . Actually may be a good thing.
Are you building for humans, or for the machines that serve them?
The future of software isn't just "AI-native" - it is Agent-Native.
We're seeing a fundamental shift from software that uses AI to enhance human experience, to platforms explicitly designed for autonomous agents.
Companies like Attio, with their CRM primitives, and the emerging concept of "Email for Agents" (think AgentMail (YC S25)) are the micro-trend. It is about APIs as the product, ontologies as the moat, and domain primitives as first-class citizens.
In this latest article we explore why this distinction matters and what defines truly Agent-Native software: medium.com/@balajibal/beyond…
balaji bal retweeted
JUST IN: OpenAI’s internal projections reportedly show the $122,000,000,000.00 they raised today gives them as little as 18 months of operational runway before they need to raise again.
If you are always shouting, eventually no one can hear you.
First, there was Delve, a startup with $26 million in funding and the backing of the full Y Combinator network, responding to fraud claims with a statement an intern could probably have written better.
Then came Delve’s auditors, Accorp Partners—who insist they are not—posting on LinkedIn about the accusations, while spending some of their response blaming the anonymous whistleblower destroying the credibility of their post.
Now Anthropic has added insult to self-inflicted injury, sending out DMCA requests long after the horse has bolted, after its own staff inadvertently leaked source code. This is a company that should probably be careful about making copyright claims, given how often it has positioned itself against them.
The best thing would probably have been to let it quietly slide under the carpet.
All of this makes me think Silicon Valley PR has become little more than tall claims and loud talk. And when that is all you can do, you lose all nuance.
balaji bal retweeted
for added context: when a 17yo developer (@xiaonweb) politely pointed out that bragging about LOC is silly, Garry's response was to publicly call them a "clout farmer."
the "clout farming" teen... wrote a browser engine in Rust at 17. HTML tokenizer, CSS cascade, box model layout, GPU renderer via wgpu, and published a technical breakdown showing deeper understanding of how the web works than most senior engineers I've worked with (including me, cos I've never dug that low-level).
vs
the "shipping" guy, the president of Y Combinator, a multi-billion dollar startup kingmaker, who mass-generates code with 113 Claude sessions a week, counts lines like a Duolingo streak, and ships test files, 0-byte AVIFs, and 4 MB uncompressed PNGs to production.
right, punch down at a teenager. on main @.
xitter.cf/xiaonweb/status/203862…
