标签：#ai

The same agent, same prompts, did markedly different work on two codebases I work in. One has a test suite that runs in eight seconds. The other takes twelve minutes. The eight-second project gets a careful, iterative collaborator. The twelve-minute project gets a confident guesser. I noticed it first as a vibe. The agent in the slow codebase would write five files at once, then announce the task complete without having run anything end to end. The agent in the fast codebase would write one function, run the tests, react to the failure, fix it, run them again. Same model. Same configuration. The only difference was how expensive it was to learn whether the previous step was right. That is the whole post in one sentence. An agent's effective intelligence is bounded by how fast it can verify its hypotheses. Cut the verification cost and you raise the agent's apparent IQ. Raise it and you lower the agent's apparent IQ. The model in the middle is unchanged. Why this binds harder for agents than for humans A human engineer can hold a hypothesis in their head. "I think this works. I will check it later." The cost of holding the hypothesis is roughly free; the human has institutional memory, intuition, a sense of what the code does that does not require running the code to confirm. They can defer verification without losing fidelity. An agent cannot. It has no intuition about your codebase. The only ground truth it has access to is what the tests say, what the type checker says, what the build says. When those signals are cheap, the agent uses them constantly. When they are expensive, the agent stops using them and starts speculating. Speculation by an agent looks plausible. It produces code that compiles, follows the patterns it has seen in your repository, names things sensibly. The problem is that plausible is not the same as correct. The agent that speculates is shipping a guess; the agent that iterates is shipping a tested answer. From the diff alone, they can be hard

Everyone is adding AI to their product right now. Most of them are doing it wrong. Not because they chose the wrong model. Not because they used the wrong library. But because they treated AI integration like a regular feature and skipped all the engineering discipline that production systems require. I have integrated LLMs into multiple production applications. This is what I wish I had known before I started. The Mental Model Shift You Need First A traditional API call is deterministic. You send a request, you get a predictable response. You can write tests against it. You can cache it. You can reason about it. An LLM call is not deterministic. The same input can produce different outputs on different runs. The model can refuse, hallucinate, or return output in a format you did not expect. Your system needs to be designed around this reality, not in spite of it. This means defensive parsing, fallback logic, output validation, and graceful degradation are not optional extras. They are the core of the feature. Choosing the Right Model for the Right Job The biggest LLMs are not always the right choice. I learned this building EditDeck Pro, an AI creative platform for music. Some tasks needed a large frontier model for nuanced creative output. Others needed a fast, cheap model that could run many times per session without accumulating significant latency or cost. The pattern that works: Use a lighter model for classification, extraction, and short structured outputs. Use a larger model for generation tasks where quality matters more than speed. Route dynamically between them based on the task type. This can reduce your inference costs by 60 to 80 percent on workloads that mix simple and complex tasks. Prompt Engineering Is Software Engineering Prompts are code. They should be versioned, tested, and reviewed like code. I store prompts in a dedicated module with version numbers. When I change a prompt I run it against a fixed evaluation set of inputs and compare the out

I was looking at my bank statement one day and realised I was paying for 4 things I completely forgot about. Combined it was around 40 euros a month just silently leaving my account. I'm a 17 year old developer from Cyprus and I spent the last few weeks building Capsule, a simple subscription tracker that shows you everything you pay for, alerts you before renewals, and tracks how much you save by cancelling things. No bank connection required. You just add your subscriptions manually. Privacy first. It's not on the Play Store yet but the waitlist is live at capsule.crickdevs.com if anyone wants early access. Would genuinely love feedback from real people before I launch.

AI workflow automation looks simple in demos. A form submission comes in. An AI model reads it. The CRM gets updated. A Slack message goes out. An email is sent. But once you move from demo to production, the workflow becomes more sensitive. What happens if the AI summary is wrong? What happens if the CRM is updated with incomplete data? What happens if the customer request needs human approval before the next step? What happens when a workflow fails halfway? That is where AI workflow automation needs better architecture. In one recent project, we designed an AI workflow automation system using: Make.com for workflow orchestration FastAPI for custom backend logic OpenAI/GPT APIs for summarization and structured output Monday.com CRM for record management Slack for internal notifications Gmail for email-based communication Human review steps for approval and control The goal was not to build a chatbot. The goal was to reduce repetitive manual review work while keeping the workflow controlled, traceable, and practical for daily business use. The workflow problem The original workflow had several manual steps: A new request came in. Someone reviewed the request manually. Important information was extracted. A CRM record was created or updated. The internal team was notified. A follow-up email was prepared or sent. The team tracked the workflow manually. This kind of workflow is common in service businesses, operations teams, sales teams, and CRM-heavy processes. The pain was not that any one step was too difficult. The pain was that the same steps repeated again and again. That makes the workflow slow, inconsistent, and dependent on manual copy-paste work. Why not fully automate everything? The obvious idea is: Let AI read the request and update everything automatically. But that can be risky. AI-generated output can be incomplete, overconfident, or slightly wrong. That may be acceptable if the output is only a draft. It is not acceptable if the output directly updates

Large exchanges are designing derivative products around AI tokens, which are increasingly being considered less a computational output and more a raw material input, like electricity or bandwidth.

As Apple tries to shrink Gemini for the iPhone, a cloud component is probably inevitable.

CEO Ariel Katz argues that while AI can replicate workflow SaaS, it can't copy H1’s unique doctor data.

The most memorable part of 007 First Light is something that's typically pretty boring: the tutorial. In many games, you're forced through a series of tedious lessons in how to play, presented in a way that feels disconnected from the story itself and at a plodding pace. But First Light does something different. Because the […]

StrictlyVC Los Angeles is on June 18. Join for meaningful networking and fireside chats with leaders from Mach Industries, Shinkei Systems, and more. Register today.

Here’s why Anthropic and OpenAI are on board with Illinois safety testing.

The new Opus model comes with a tool called Dynamic Workflows, for coordinating swarms of subagents.

Anthropic is releasing Claude Opus 4.8 on Thursday, and the company is touting the model's "honesty." According to Anthropic, it trains "all [its] models to be honest - for instance, to avoid making claims that they can't support." But it notes that "a general problem with AI models is that they sometimes jump to conclusions, […]

Next month's Tribeca Festival will include the premiere of an AI-generated film: Dreams of Violets. The 75-minute film is a fictional dramatization of the Iranian government's mass killing of protestors in January, with the people and images fully created by AI, as reported earlier by The Hollywood Reporter. Dreams of Violets cost $2,000 to make […]

New features coming to YouTube could make it better for listening to podcasts, rolling out to Premium subscribers starting today on Android and coming later to iOS. A new "on-the-go mode" shifts YouTube into an audio-first layout, with larger, simplified playback buttons, a still image in place of the video, and a timeline showing video […]

Did you know that a 35-billion-parameter model can generate tokens at the same compute cost as a 4B model? That single fact made me abandon a multi-model agent architecture I'd spent a weekend building. But I had to run the benchmarks first to understand why. Here's the full breakdown, with commands, numbers, and the architectural reason it all falls apart on shared-memory hardware. The Discovery That Changed Everything I'd been running qwen3.6:35b on my Minisforum UM790Pro for weeks as my daily coding assistant. 17.8 tokens/second -- genuinely usable for interactive work. But I kept wondering: could I run a lightweight sidecar model alongside it for quick classification and tool-calling in an agent pipeline? Before I even started benchmarking, I dug into what qwen3.6:35b actually is under the hood. It's a Mixture of Experts model: 256 total experts with only 8 activated per token. The architecture also incorporates SSM (State Space Model) components alongside traditional attention -- Mamba-style layers that handle certain sequence patterns more efficiently than pure transformers. The math hit me: 8 out of 256 experts means each token only touches roughly 4-5B parameters worth of compute. The model carries 36 billion parameters of knowledge , but its per-token cost is comparable to a small dense model. I was planning to run a separate 4B model for "fast tasks" next to a model that already operates at 4B-class speed. But I had to prove it with numbers. Hardware and Ollama Setup The UM790Pro specs that matter for this experiment: CPU: AMD Ryzen 9 7940HS (Zen 4, 8C/16T) iGPU: AMD Radeon 780M (12 RDNA 3 compute units) RAM: 96 GB DDR5-5600 (~80 GB/s bandwidth) GPU memory pool: 2 GB dedicated VRAM + 46 GB GTT = 48 GB GPU-accessible That 48 GB GPU pool sounds enormous until you realize it's carved from the same DDR5 that the CPU also uses. There is no separate GDDR6 bus. Everything -- CPU inference, GPU inference, KV caches, OS operations -- flows through one 80 GB/s pipe.

OpenSparrow v2.6 is out. This one's a big step forward — RAG (Retrieval-Augmented Generation) integration, bulk grid operations, and a whole new UX layer. RAG & AI integration You can now upload documents and let users ask questions against them. The system retrieves relevant sections and generates answers using an LLM (supports Ollama locally or any OpenAI-compatible API). What's new: RAG Statistics tab in admin — tracks query tokens, response times, document matches, and recent queries Multilingual auto-response — user questions are answered in their own UI language, no schema changes needed 20-language support — "Ask AI" panel fully translated, plus language dropdown in the test interface Good for things like: knowledge base Q&A, customer support automation, or letting internal teams ask questions about their own data. Grid & bulk operations Mass Edit module — select rows via checkboxes, bulk edit fields, change owners, duplicate, or delete with one click Keyboard shortcuts — arrow keys to navigate, Tab, Ctrl+C to copy, Ctrl+F to search, Ctrl-hold for help modal — works across all 20 languages Quick Data Cleanup toolbar — find & replace with live preview, case sensitivity, accent-ignore. Editor-gated with audit trail. Admin improvements Renamed "RAG Knowledge Base" to "Centrum AI" (heading is translatable) Migration Manager — tracks pending cleanup tasks after version upgrades, with automatic backups and audit trail FK columns render as dropdowns in forms by default Security & Quality All bulk operations (mass edit, cleanup, delete) are editor-role gated CSRF protection on every operation Full audit trail — every change is logged 20 languages fully supported across all new modules Fixed regressions in RAG API and CSV import Following this series? OpenSparrow v2.3 – visual admin panel, zero dependencies, now with ERD and M2M support OpenSparrow – open-source admin panel builder, zero dependencies, v2.1 just dropped Websites opensparrow.org github.com/wrobeltomasz/

One poisoned extension, one package install, one CI workflow. Any of them can now be the first domino. That is the uncomfortable lesson from the latest Shai-Hulud activity and GitHub’s recently confirmed internal-repository breach. The scary part is not only the number of affected packages, tokens, or repositories. Counts move fast. The scarier part is where the attacker code ran: inside the trusted developer and CI path. The modern supply chain is not just “the dependencies we ship to production.” It is your IDE, your package manager, your GitHub Actions runner, your cache keys, your OIDC flow, your local gh auth, your AI coding tool config, and the cloud account that quietly pays the bill when something goes sideways. What happened, briefly CISA described the original Shai-Hulud wave as a self-replicating npm worm that compromised more than 500 packages and targeted GitHub personal access tokens plus AWS, GCP, and Azure keys. GitHub later said it removed 500+ compromised packages and began pushing npm toward shorter-lived credentials, 2FA enforcement, and trusted publishing. The later waves got more CI-aware. Instead of only stealing npm tokens from maintainers, they looked for credentials inside build environments, abused publishing workflows, and used the build system itself as distribution. Microsoft’s May 2026 reporting on the @antv ecosystem described a “Mini Shai-Hulud” style campaign that targeted GitHub Actions environments and stole GitHub, AWS, Vault, npm, Kubernetes, and 1Password secrets. Microsoft said GitHub removed 640 malicious packages and invalidated 61,274 npm granular access tokens with write permissions and 2FA bypass. Then GitHub confirmed an incident involving a compromised employee device and a poisoned third-party VS Code extension. GitHub said the attacker’s claim of roughly 3,800 internal repositories was “directionally consistent” with its investigation, while also saying its current assessment was exfiltration of GitHub-internal reposi

In 1994, Peter Deutsch published the Fallacies of Distributed Computing — eight assumptions that every developer building distributed systems makes, discovers are wrong, and pays for in production. The network is reliable. Latency is zero. Bandwidth is infinite. Each assumption sounds true. Each leads to system failures that could have been avoided. Thirty years later, we're making the same category of mistakes with generative AI. The trough of disillusionment for AI-assisted development has begun. Byron Cook, VP and Distinguished Scientist at Amazon, founder of AWS's Automated Reasoning Group (300+ scientists, 15+ teams), says it plainly: "Generative AI is sliding into the trough of disillusionment." The headlines are shifting. The "summer of vibe coding" is over. The disillusionment isn't caused by AI being useless. AI-assisted coding delivers real productivity gains. The disillusionment is caused by false assumptions about WHERE the gains come from and WHAT changes when generation gets fast. Teams expected 10x engineering. They got 10x code generation and 1x everything else. The gap between expectation and reality is the trough. This series names the eight assumptions, explains why each one fails, and presents the resolution — not from theory, but from domains that hit the same wall and climbed out. The Eight Fallacies 1. Faster code generation means faster engineering. You made one sub-system 10x faster. Seven others didn't change. The system doesn't get faster — it breaks at the interfaces. The CPU-memory wall tells you exactly what happens and what fixes it. 2. If the output looks correct, it is correct. AI-generated code is optimized for plausibility, not correctness. It compiles, passes tests, and reads well — while violating properties nobody tested. Plausible is not correct. The gap is where production failures live. 3. You can verify AI output with another AI. Guardrails, LLM-as-judge, AI code review — the verifier has the same failure modes as the thing

Every week another "AI agent for X" launches. Email triage. Calendar coordination. Sales follow-up. PR reviewer. Slack monitor. Meeting summarizer. I've installed enough of them to see the pattern. Here's the dirty secret nobody mentions in the launch posts: These tools don't reduce your work. They multiply your notifications. Each AI tool is configured to be helpful by default. "Helpful" means: "I noticed this thing — here's a notification." Stack a dozen of those, and instead of one inbox to ignore you have twelve. The signal-to-noise ratio gets worse every time you add an AI to your workflow. The mainstream answer is "just configure each one." Sure. Spend four hours tuning notification settings every time you add a tool, and another four hours when one of them ships a "smarter notifications" update. That's not productivity. That's notification janitorial work disguised as setup. This is a structural problem. Not a configuration problem. The wrong question Every AI tool asks the same thing: "Is this important?" Wrong question. There is no objective "important." Importance depends on you, right now. A Stripe webhook is important when you're debugging a checkout flow. The same webhook is pure noise during a deep work block. A Slack message from your cofounder is critical at 11am Tuesday and irrelevant at 11pm Friday. The right question is: Is this urgent enough to interrupt me, right now, given what I'm doing? That's not a question any individual AI agent can answer. It's a layer above all your AI agents. None of them have the context. None of them know what the others are doing. None of them know how you're spending the next hour. So they all default to "I'll just send you a notification, you decide." Which is exactly the experience you have right now: drowning. What an AI firewall actually looks like I'm building that layer. It's called Klorn . Here's how it works in practice. Every signal — email, calendar invite, agent action, webhook, push from another tool — g

By Umair Sheikh, founder of Gateplex Autonomous AI agents are shipping fast. LangChain, CrewAI, AutoGen — the frameworks are mature, the tutorials are everywhere, and developers are connecting agents to real systems: databases, payment APIs, email, file storage. And then something goes wrong. Not because the code is buggy. Because the agent did exactly what it was told — and what it was told turned out to be a problem nobody anticipated. I spent nearly a decade in fintech and responsible AI policy watching this pattern repeat. A system behaves perfectly in testing. In production, an edge case triggers behaviour that was technically correct but operationally catastrophic. By the time anyone notices, the action has already executed. The problem is not the agent. The problem is that there is nothing between the agent's decision and the real world. The gap nobody talks about Most agent observability tools log what happened. That is useful for debugging. It does nothing to prevent the next incident. What agents actually need is a governance layer — something that intercepts every action before it executes, checks it against your rules, and either allows it, flags it for review, or blocks it outright. This is what a firewall does for network traffic. Your AI agent deserves the same treatment. What this looks like in practice Here is a simple LangChain agent calling an external tool: from langchain.agents import initialize_agent , Tool from langchain.llms import OpenAI def send_payment ( amount : str ) -> str : # This actually moves money return f " Payment of { amount } sent " tools = [ Tool ( name = " SendPayment " , func = send_payment , description = " Send a payment " )] agent = initialize_agent ( tools , OpenAI (), agent = " zero-shot-react-description " ) agent . run ( " Send $5000 to vendor account " ) This works. It also has no guardrails whatsoever. If the agent misreads the input, hallucinates a vendor, or gets manipulated via prompt injection, the payment goes