LTX 2.5

Tue, 20 Oct 2026 00:00:00 GMT

Official page for LTX Video 2.5 — an open-source AI video generation model by Lightricks. Optimized for high-resolution video creation, realistic motion, and fast inference directly from text or image prompts.

Permalink

Note: Anatomy: the object is the explanation Most technical diagrams are boxes and…

Sun, 18 Oct 2026 07:32:00 GMT

Anatomy: the object is the explanation

Most technical diagrams are boxes and arrows. They show that the parts exist, but not what the parts do — you have to imagine the motion yourself.

Anatomy takes the opposite approach. It’s a skill for Claude Code that, given one line naming an idea, invents a real machine whose mechanism maps one-to-one onto that idea, then draws it as a live isometric figure. A queue becomes a marble run. An LRU cache becomes a bookcase with four places on top. A bloom filter becomes a fruit machine whose three reels are three hash functions. The author describes it in a single rule: the object is the explanation.

Each figure is built from 40 to 500+ solid parts, shaded with a single light source and one highlighted “watch this” part. You can drag, hover and use the keyboard; there’s an idle tour, it pauses when off-screen, and it goes still under reduced motion. Every readout is computed from the same model that animates the parts, so the picture and the numbers can’t disagree.

Figures ship as framework-free SVG/TypeScript or React components (anatomy-figures), with five themes, and can also be exported as printable A4 sheets. It’s MIT-licensed.

Quote by Satya Nadella

Thu, 08 Oct 2026 18:40:00 GMT

In the future, every company will be a software company.

— Satya Nadella

AI Agents, Clearly Explained

Thu, 01 Jan 2026 00:00:00 GMT

Imagine hiring a new personal assistant. You don’t hand them a single tiny task. You say: “Plan and book my summer vacation to Italy — within my budget, my love of art and pasta, and my wife’s preference for coastal views. Keep me posted, and only ask me questions when it matters.”

A good assistant breaks that into dozens of steps: research flights, compare hotels, draft an itinerary, make bookings, sync it to your calendar. You gave one sentence. They returned a finished plan.

That is the essence of an AI agent. Not a tool that answers a prompt, but a system that takes a goal, works out the steps itself, and keeps going until the job is done.

The short version: a chatbot responds; an agent acts, checks the result, and decides what to do next.

From reactive to agentic

For years, AI has been excellent at pattern recognition — spotting cats in photos, translating languages, writing passable prose. But almost all of it has been reactive. You ask, it answers. You input, it outputs. The intelligence is real; the initiative is not.

Agents change the unit of interaction. Instead of a prompt, you hand over an objective. Instead of a reply, you get a completed task — or a well-argued explanation of why it can’t be completed.

What exactly is an AI agent?

At its core, an agent is a system that perceives its environment, makes decisions, and takes action to reach a goal, often without supervision.

Concretely, it runs a loop:

Step	What happens	In the Italy example
Perceive	Takes in the goal and any new information: a result, an error message, a page it just read.	Reads your budget and preferences, then loads current flight prices.
Plan	Breaks the goal into steps and chooses the next one.	“Flights first — hotels depend on the dates.”
Act	Uses a tool to do something: search, run code, call an API, write a file.	Queries a booking API for Milan–Rome options.
Observe	Reads what came back, including failures.	“That flight is sold out. Three alternatives within budget.”
Repeat	Loops until the goal is met or it needs your input.	Revises the plan, books the hotel, asks about the coast.

The important part is the loop. A single model call cannot be an agent, no matter how good the model is, because there is nothing to loop back to. Agency is a property of the system, not the model.

The coffee maker and the barista

A chat model on its own is like a very good espresso machine. Press “cappuccino” and you get a perfect cappuccino every time. One input, one output. It will never decide you might prefer a latte today.

An agent is the barista. You say, “I need something to wake me up, but I don’t like bitter coffee.” They read the request, weigh options, ask whether you’ve tried a particular roast, then use their tools — grinder, machine, steamer — to make and hand you the drink. Multi-step, tool-using, adapted to your actual goal.

The analogy isn’t perfect, and it’s worth naming why: many chat products now bolt agent features onto the chat window, which blurs the line. The distinction still holds underneath — the difference is whether the system chooses and executes steps, or merely produces text.

A quick test

Three questions separate a real agent from a chatbot wearing the label:

Does it choose its own steps? Or does your prompt spell out every action?
Can it change something outside the conversation? Sending an email, writing a file, booking a room — actions with consequences.
Does it react to what happened? If step three fails, does it try step four differently, or does it stop?

If you answer no to all three, you have a chat interface, and that may be exactly what you need.

The building blocks

Most capable agents share four components.

The brain (a language model). The LLM is the reasoning engine, not just a text generator. It interprets the goal, drafts the plan, and picks the next action. Everything else in the stack exists to give that reasoning something to work with.

Planning and reasoning. This is where agents earn their name. Instead of one straight shot, the agent drafts a to-do list, weighs alternatives (“flights first, or check hotel availability?”), and corrects course when something fails. Some agents plan explicitly and re-plan as they go; others decide one step at a time. Either way, the plan is provisional, not a script.

Tools (the hands). An agent’s reach is set by what it can touch:

Web search and browsing for current information
A code interpreter to calculate, analyse data, or manipulate files
APIs and software to send email, update a spreadsheet, book a ticket, control a device

A model with no tools can only think. A model with tools can act.

Memory. Short-term memory holds the context of the current task — what’s been tried, what came back. Long-term memory persists across sessions: your preferences, past outcomes, the corrections you’ve already given. Long-term memory is what turns a competent stranger into a useful assistant, and it is also where most privacy questions live.

What this looks like in practice

Today, reliably:

Research assistant. “Find the five most recent papers on climate impacts on European wheat yields, summarise where the authors disagree, and save it to a Google Doc.” The agent searches, reads PDFs, synthesises, writes the file.
Shopping researcher. “Find waterproof winter coats under $200 with strong durability reviews. Send me three options by Friday.” It searches, filters, compares, reports.
Coding agent. “This script fails to connect to the database. Fix it and explain what you changed.” It reads the traceback, inspects the code, runs tests, commits a fix.

On the horizon, harder:

Business process. “Onboard John Doe: email, licences, payroll, orientation meetings.” Requires touching HR, IT, and finance systems at once.
Care coordination. Monitoring a recovery plan across medication schedules, wearable data, and appointments. The technical work is tractable; the stakes and the liability are the hard part.
Scientific discovery. Given a target, an agent hypothesises, designs virtual experiments, runs simulations, and proposes the next wet-lab test.

Why it’s a big deal

It changes what AI is for. The value shifts from producing text to finishing work. That is a different product category, not a better version of the old one.

It spreads expertise. You no longer need to be a data analyst, a travel agent, or a legal researcher to get those skills applied to your problem. The tooling is borrowed.

It compounds. Agents automate multi-step cognitive workflows, not just single repetitive actions. Hiring a tireless researcher who never loses the thread is a different proposition from buying a faster calculator.

It’s a step toward more general systems — with an asterisk. Agents that set sub-goals, learn from feedback, and sustain action over long horizons are clearly moving in that direction. But capable agency is not the same as general intelligence, and the gap is larger than the enthusiasm suggests.

Where it goes wrong

Reliability. An agent can build a plausible plan on a hallucinated fact and execute it flawlessly. Confident, coherent, and wrong is the default failure mode. Verification is the real bottleneck — not generation.

Cost and speed. Every loop iteration can mean another model call and another API request. A task that takes a person two minutes may take an agent thirty calls. It’s easy to build something impressive that is too slow or too expensive to run.

Safety and control. How do you bound what an agent may do? The alignment problem stops being abstract the moment your agent can spend money or send email on your behalf.

Security. An agent holding credentials to your inbox and bank account is a powerful thing to compromise. Prompt injection — instructions hidden in a webpage or document that the agent reads and obeys — is an unsolved practical problem, not a theoretical one.

The work question. Like every automation wave, agents will reshape jobs: some tasks absorbed, some roles amplified, some displaced. The honest answer is that the mix is unclear and the speed matters more than the direction.

When not to use one

Agents are powerful and frequently the wrong tool. Skip them when:

The task is one step. A single API call or database query doesn’t need a reasoning loop.
The answer must be deterministic. If you need the same output every time, write code, not an agent.
Latency matters. Multi-step loops are slow by construction.
The action is irreversible and high-stakes. Sending money, deleting records, filing documents — an agent can prepare those, but a person should pull the trigger.

The best agentic systems are narrow, tool-rich, and supervised: plenty of capability, clear boundaries, and a human at the decision points that matter.

From a chat box to a goal box

We’re moving from the era of the single prompt to the era of the stated objective. The interface of the future may not be a text field you type into, but a goal you assign and monitor.

That changes what makes someone effective. The people who thrive won’t be those with the most clever prompts, but those who can:

State goals precisely — outcomes, constraints, and what “done” looks like
Judge results critically — spot the confident, fluent, wrong answer
Delegate deliberately — know which decisions to hand over and which to keep

Which is to say: the same skills as managing a good team. The digital assistants are getting better quickly. The useful question is no longer what they can do, but what you’d actually hand them.

Fuat Beşer