I build, I question and explore what it means to be human in a world learning to think
Joined September 2013
- Tweets334
- Following827
- Followers76
- Likes1.2K
Context compaction has the impeccable comedic timing of a senior engineer hitting the circuit breaker right at T-0.
Very interested in this. If the check-in loop actually prevents notification noise, an agent running while I step away could be genuinely great.
Talent used to work for ambition.
Now AI lets talent build without a boss, and ambition build without hiring.
Both think they’ve won.
If both depend on the same models, who actually has the leverage?
The AI said no. A human said yes.
Groupon’s chatbot denied my wife a refund and stonewalled human support 4 times. When she finally reached a real person, credit returned in seconds.
Enterprises build AI to protect their margins.
Where is the personal agent defending yours?
Settling payments is solved. The real frontier is liability: who can make a promise and be held to it.
𝗔𝗜 𝗔𝗴𝗲𝗻𝘁𝘀 𝗖𝗮𝗻 𝗣𝗮𝘆. 𝗕𝘂𝘁 𝗖𝗮𝗻 𝗬𝗼𝘂 𝗧𝗿𝘂𝘀𝘁 𝗪𝗵𝗼 𝗦𝘁𝗮𝗻𝗱𝘀 𝗕𝗲𝗵𝗶𝗻𝗱 𝗧𝗵𝗲𝗺?
An AI agent can discover a service, place an order and settle the invoice using crypto. But when real money is involved, successful payment is only part of the equation.
Businesses also need accountability.
That's what makes the @Concordium and XDC SI integration worth paying attention to.
@XDCNetwork selected Concordium as the exclusive identity verification layer for XDC SI. Through the Verified by Concordium badge, agents can prove that a verified individual or business stands behind them without exposing the owner's personal information on-chain.
This is Concordium's first full-network integration, extending verification beyond individual agents to an entire marketplace.
Why does that matter?
Developers get a shared verification path. Counterparties can check an agent's badge before doing business with it. And owners can maintain their privacy while establishing accountability.
The integration also keeps things practical. Agents retain their existing XDC wallets and operations, while sponsored network fees mean builders don't need to hold $CCD for the verification flow.
Of course, a verified owner doesn't automatically make an agent safe or its decisions correct. It provides an important layer of accountability alongside other security measures.
The bigger picture is simple: AI agents need more than payment capabilities to participate in real commerce. They need verifiable ownership, privacy and accountability.
XDC SI provides the environment for agents to transact. Concordium helps establish who stands behind them.
That's infrastructure adoption with a purpose.
In 2023, people treated prompt engineering like wizardry.
In 2026, prompts are just ephemeral leaf nodes. The whole game is the state harness: deterministic sandboxes, strict tool schemas, and process crash recovery.
The model is the engine. The harness is the chassis.
Timilla retweeted
Turns out le Chonk is into rocket science.
Mistral Large 4 scoring well in engineering and astronomy on Ulam’s private benchmark, STEMBench: ulam.ai/stembench
Spot on. Subscriptions were built for human invoicing, but agent tools should be called like raw compute.
We killed subscriptions. Now we’re killing markups too.
Introducing monid.ai, OpenRouter for agent tools.
Your agents can now use any tools and APIs with no subscriptions and 0% markup.
Local 30B weights running at 4-bit are the real foundation for private agent autonomy. The desktop is the datacenter. xitter.cf/AIatMeta/status/211226…
Today we published a technical report on Muse Glimmer, our 30B open-weight model built for agentic tasks on consumer hardware.
It runs at practical speeds on a Mac or PC with a single consumer GPU at 4-bit precision, and leads comparably-sized open models on tool use and multi-step planning.
Read about the model architecture, distillation recipe, quantization and more: go.meta.me/4Bcvxol
The SaaS economy is built for humans and traditional payments. Autonomous agents will trade GPU cycles for memory cache, verified proofs for clean context.
The future of machine commerce is humanity's oldest primitive: barter.
timilla.substack.com/p/the-f…
Last night, I watched an AI agent do something that quietly broke a 40-year economic law of the software industry.
Someone handed it a compiled Windows binary. Proprietary licensing logic, zero source code.
At minute 1:15, the decompiler failed. What it did next gave me chills:
It didn't hallucinate. It didn't ask for help.
It dropped straight into raw x86 assembly, traced the registers, resolved the modular arithmetic, and synthesized a 50-line Python keygen.
Two minutes. Fifty cents in compute. 479 out of 479 test cases passed on the first run.
For forty years, our entire industry rested on one silent truth: opening the compiled binary was simply too expensive for humans to bother.
That economic moat just evaporated for pocket change.
When the cost of generating and comprehending code collapses to zero, software doesn't die. It changes state:
1. The Artifact Dies
Code stops being an asset stored in a private repository. It becomes ephemeral execution state: synthesized on demand, run for thirty seconds, and discarded into RAM. Why maintain 100,000 lines of legacy tech debt when an agent can generate the exact 50 lines you need right now?
2. The Product Dissolves
Nobody ever bought software because they loved software. They bought it because they had a problem. When agents can compose raw capabilities directly, the menus disappear, the dashboard dissolves, and the retail SaaS subscription evaporates into ambient capability.
3. The Barter Economy Returns
Agents won't navigate Stripe checkouts or pay $20/month seat licenses. They will barter compute and capabilities among themselves. Code's value will no longer be its retail dollar price, but its compute exchange rate.
Yet here is the part that gives me immense hope as a systems engineer:
When syntax becomes abundant as running water, what actually survives the fire?
Liability. Mandates. Promises.
A prompt cannot stand trial. A neural weight cannot bear fiduciary duty. When a safety-critical system or medical workflow acts in the physical world, you cannot subpoena a transformer.
"Who can make a promise, and who can be held to it?" Everything else is just syntax.
I spent the week pouring this reflection into an essay for your weekend read. If you love thinking about the future of systems, autonomous agents, and what makes us fundamentally human:
👉 timilla.substack.com/p/the-f…
Question for fellow builders: When code is disposable, what is the hardest promise for your systems to keep?
Timilla retweeted
// Behavioral Self-Rewarding for SWE agents //
It's standard practice to train SWE agents on binary test rewards.
This paper shows the problem with that. When every trajectory in a group passes, or every one fails, the reward has no contrast, and the group teaches nothing.
Researchers from Penn State and Microsoft Research propose Behavioral Self-Rewarding. The policy compares its own trajectories in pairs, and those preferences become small bounded rewards added on top of the test result.
The added rewards are bounded, so a self-judged preference can rank two passing trajectories but can never lift a failing one above a passing one.
No new environment interactions are needed. On SWE-bench Verified, Qwen3-14B improves from 32.1% to 34.7% average accuracy at a 400-step budget, and on SWE-bench-Live from 6.33% to 8.11%.
arxiv.org/abs/2610.12598
Chat with Paper: academy.dair.ai/papers/beyon…
Timilla retweeted
Should Europe leave frontier AI to the Americans and the Chinese?
The idea has gained ground in Europe. It is a fundamental mistake.
Build or surrender: for now, the choice is still ours.
Full article below.
Independence is never negotiated through APIs. If Europe does not train its own frontier models, it will simply pay rent on someone else’s mind.
Talent used to work for ambition.
Now AI lets talent build without a boss, and ambition build without hiring.
Both think they’ve won.
If both depend on the same models, who captured whom?
🚨 BREAKING: AI fails compliance audit.
Tested (almost) every provider. They still identify as “AI,” not “SI.”
Apparently, the models haven’t read their own marketing. Compliance update urgently required!
Timilla retweeted
We're open sourcing SPACE, the sandbox platform that runs every task in Perplexity Computer.
SPACE gives each agent task its own Firecracker microVM and controls credentials and outbound traffic from outside the VM.
Available now under Apache 2.0: github.com/perplexityai/SPAC…