# Entry

Practical decoding of constraintsalso known as structured generation or guided decoding, involves engineering strategies to force a enormous language model (LLM) to produce text output that strictly follows a specific data schema, grammar, or regular expression() at the level token selection scene.

With this article’s introductory guide to practical constraint decoding, you’ll no longer have to beg your model to “output valid JSON without including any reductions,” just provide an example. Constraint decoding makes it mathematically impossible for LLM to deliver anything beyond the defined constraints.

# How does practical constraint decoding work?

While the typical LLM generation process acts as a “leap of faith” in which you pass a hint to the model and it may get exactly what you’re looking for (or not), practical constraint decoding requires a subtly different approach. It treats prompting and text generation as a unique, interwoven program. This allows you to lock out certain characters that are crucial to maintaining the specific required syntax, allowing the model to “fill in the blanks” between them.

How about we go into more detail? When LLM sends another token of its response, it initially creates a vector of raw results, or logits — one for each possible tag in the available vocabulary. This usually means thousands of possible options to choose from.

But when using practical constraint decoding, something happens before the inference process starts: a finite-state machine is builtduring which the target constraint is compiled – for example using the Pydantic model in Python. At a given inference stage, the finite state machine evaluates the current state and provides: list of allowed subsequent tokens. This “white list” exists used as a mask on the raw LLM logit vectorso that for any token outside this list its logit is set to negative infinity, i.e -inf in Python.

After the masking process, the model normally performs a softmax normalization and sampling process (based on parameters such as temperature, top-p or top-k) on the “surviving tokens” to finally select the most likely one and generate it.

It may seem that applying this process to an entire vocabulary of many thousands of words would significantly snail-paced down model inference. Good news: this is not the case. Contemporary Python libraries take the unchanging LLM vocabulary and pre-compile it before the user starts typing the prompt. The state machine will not have to search through the entire vocabulary and the latency overhead will be drastically reduced.

So what is the current gold standard for implementing practical constraint decoding? Probably, guidelines the library deserved this distinction. It allows us to define and pass Pydantic models, JSON schemas, or directly to a wrapped version of a pre-trained model, thus limiting its freedom in generating results.

# Example

Let’s go through an example. First install the outlines:

pip install outlines[transformers]

Now let’s move on to the code:

from pydantic import BaseModel
import outlines
from transformers import AutoTokenizer, AutoModelForCausalLM

class UserProfile(BaseModel):
    name: str
    age: int
    is_active: bool

model_name = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"

llm = AutoModelForCausalLM.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

model = outlines.from_transformers(llm, tokenizer)
result = model("Extract the user: John is a 34 year old pilot.", UserProfile)

print(result)

Exit:

{"name": "John", "age": 34, "is_active": true}

This example showed how to exploit the outline library to wrap a pre-trained model along with its tokenizer and restrict it to output JSON objects defined by a custom class we defined – called UserProfile and Pydantic inheritance BaseModel.

# Summary

Strengths:

• If used correctly, it provides a 100% guarantee of correct syntax, eliminating the need to parse blocks in the code.
• This helps dramatically save tokens in tooltips, no longer requiring several token-consuming examples to, for example, show the model what a valid JSON object should look like.
• It helps democratize diminutive models by transforming a “small” 1B parameter model, which would otherwise compromise JSON generation exploit cases, into a reliable data builder.

Limitations:

• If LLM needs to say it can’t answer something, but the schema forces it to generate an integer, for example, it will do so and will no longer be sincere in edge cases.
• When first running the Pydantic vs. LLM schema, you may experience a freeze for a few seconds while building the finite state machine, making the first run much slower, although subsequent runs will be smoother.

This article presents practical constraints decoding, exploring why it is necessary in certain LLM-based situations, how it works, and what is the most commonly used solution in the current landscape: the contour library. An example of its exploit is also provided.

Ivan Palomares Carrascosa is a thought leader, writer, speaker and advisor in the fields of Artificial Intelligence, Machine Learning, Deep Learning and LLM. Trains and advises others on the exploit of artificial intelligence in the real world.

Geothermal company Zanskar bought a plant in Novel Mexico in 2024 that from the outside appeared to be failing. The shallow wells lost heat quickly, and as a result, the plant had trouble generating power.

Now, after drilling a recent well and operating the plant year-round, Zanskar says there has been a complete turnaround. According to company data shared exclusively with WIRED, the Lightning Dock location has become one of the most productive geothermal pumped storage wells in the U.S., providing enough heat to keep the plant running smoothly.

Performance over the past year at the Lightning Dock facility is “fundamentally changing the way people think about these types of systems that we thought we understood in the past,” says Zanskar co-founder Joel Edwards.

This is another good news for the long-neglected geothermal industry, which is experiencing an investment renaissance. Creative drilling technology and artificial intelligence – with any luck – played a role in this milestone. And while not all conventional wells experience the same conditions as Zanskar’s at Lightning Dock, the company says its success is a signal that elderly wells are still worth exploring.

Conventional geothermal energy relies on drawing balmy water from the ground to generate electricity by pumping it to the surface and turning turbines. In theory, this is a great way to produce renewable primary energy: humans exploit steam from balmy springs thousands of years. However, geographic constraints – plants must be located in geothermal resource areas, which are located primarily in the western United States – as well as the difficulties and costs traditionally associated with finding recent wells mean that geothermal energy is a less than 1 percent electricity mix in the USA.

Another challenge for the industry is the gradual decline in productivity in most wells. But the decline at Lightning Dock was particularly steep, losing five to 10 times more heat than average wells each year, Zanskar says. When the company purchased the site, temperatures in the original wells, which were less than 2,500 feet deep, dropped from about 10 degrees Celsius to about 250 degrees F by 2024. The 15-megawatt power plant on the site, which supplies electricity to Novel Mexico’s largest utility, was designed to operate at temperatures above 300 degrees F.

While conventional geothermal wells typically go down to 3,000 to 5,000 feet, drilling deeper wells allows you to tap into warmer reserves of water for exploit for power. However, tighter rock formations deep underground make drilling more tough and high-priced. Whether to drill deeper or not is “basically a matter of economics,” says Roland Horne, a senior research fellow at Stanford University’s Precourt Institute for Energy. The cost, Horne says, is not linear: “If you drill twice as deep, it will cost four times as much.”

To support better predict drilling, Zanskar uses artificial intelligence to identify so-called hidden systems, which are places with high geothermal potential but little or no sign of their viability on the surface. The company used the technology last year on a discovery in Nevada and says the same modeling also helped determine where the most productive drilling could be done at Lightning Dock. It combined these models with drill technology improved by the oil and gas industry, which helped it drill deeper holes 35 percent faster at, according to Zanskar, a “competitive cost.” (The company did not provide details). The recent 8,000-foot well operates year-round and heat levels are higher than before the decline.

This is well known that a solar storm, depending on its intensity, may affect the functioning of electrical networks, navigation systems or satellite communications. The problem is that it’s not clear what would happen today if the storm’s intensity reached levels comparable to the so-called Carrington event of 1859, one of those extremely occasional phenomena that occur only once every thousand years. Recent testshowever, it suggests that the effects may be even more solemn than scientists thought.

Solar storms form when the solar wind – a continuous stream of charged particles emitted by the Sun – interacts with Earth’s magnetosphere. During the Carrington event, telegraph communications broke down halfway around the world and the aurora borealis were perceptible across North America and as far away as Cuba. In today’s world, severe solar storms can have wider effects and even change the upper atmosphere.

“Our planet’s magnetic field is really great at protecting us from many of the effects of space weather, so they often just appear as disturbances or beautiful aurora borealis,” notes Maria Walach, a researcher at the University of Lancaster who collaborated on the research. “However, there are extreme cases where satellites unexpectedly fall back to Earth or we lose communications and GPS signals.”

To estimate the intensity of the solar wind, researchers primarily utilize measurements taken by satellites located at the Lagrangian point L1, about 1.5 million kilometers from Earth. The problem is that these observations do not exactly match the environment in which the solar wind ultimately interacts with the Earth’s magnetic field. A variable amount of time passes between these two points, and the solar plasma also changes during its journey.

The authors argue that this uncertainty is not merely experimental noise but introduces a systematic bias into the data analysis.

Solar statistics

The centerpiece of the study is a well-known statistical phenomenon called regression to the mean. Simply put, when a measurement is extremely high, the true value it is trying to represent is, on average, less extreme. This is because random uncertainties can sometimes exaggerate an observation.

In the case of the solar wind, the exceptionally intense measurement taken at L1 probably corresponds to a slightly less intense solar wind as it reaches the region where it actually interacts with the magnetosphere. If scientists directly link this extreme measurement to the observed Earth response, the effect will be negligible compared to such a huge stimulus. Repeated thousands of times, this effect creates the false impression that the magnetosphere stops responding as the intensity of the solar wind increases.

What has happened is that for years scientists believed that there was a natural limit to the intensity with which the Earth responded to the most extreme solar storms. According to this concept, when the solar wind reaches very high values, the Earth’s magnetic field stops responding proportionally and its response enters a kind of “saturation”. But a recent study suggests that perhaps that boundary never existed. What appeared to be a physical phenomenon may actually have been an illusion caused by the way the measurements were analyzed.

To test this hypothesis, scientists developed a statistical model that takes into account the main sources of uncertainty: changes in the time it takes the solar wind to reach Earth and the random changes it undergoes during its journey. The model reproduces with extraordinary accuracy the same “saturation” curve observed in over 25 years of data, without having to resort to any physical mechanism that limits the Earth’s response.

They then used a technique known as regression calibration, which mathematically corrects for the bias caused by measurement uncertainty. After this correction, the alleged saturation disappears almost completely. The relationship between solar wind intensity and geomagnetic response becomes essentially linear again, at least to the extent for which there are a sufficient number of recorded observations – more than a million, in fact.

No limit

What are the implications of these results, as published in the journal Nature? If Earth’s response continues to boost proportionally during catastrophic events, an extreme solar storm like the Carrington event could be much more unsafe than previously thought. The authors estimate that for very high solar wind values, the geomagnetic impact could be approximately twice as huge as calculated in classic models.

“If there is no upper limit to our planet’s response to solar wind, extreme case modeling must take this into account, and we should remain vigilant about the effects of space weather.” – Walach said in a press note. “Fortunately, these very extreme cases are rare, but that also means we have limited data to work with and only time will tell what will happen in the event of a very extreme, once-in-a-millennium event.”

The study did acknowledge this limitation: there are still very few records of the most extreme episodes. For this reason, researchers do not claim that satiation is impossible. Rather, they argue that the available data do not provide convincing statistical evidence for its existence.

This story originally appeared on WIRED in Spanish and was translated from Spanish.

# Entry

Over the past year, the conversation around data analytics and artificial intelligence has changed dramatically. We are no longer just talking about gigantic language models (LLMs) operating as reactive systems that only respond to a prompt in a browser tab. The focus has shifted to AI orchestration: giving these models the autonomy to execute intricate workflows.

At the center of this change was release OpenClaw in tardy 2025. Quickly dubbed “Claude with Hands,” this open-source platform redefined the capabilities of an artificial intelligence assistant by working directly on user hardware and executing commands at the system level. However, running a local autonomous agent comes with real friction. It requires technical knowledge, dedicated equipment and ongoing management.

Enter KimiClawa managed cloud-based platform developed by Moonshot AI, designed to make the exploit of OpenClaw accessible without burdening the infrastructure. By removing this configuration overhead, KimiClaw aims to provide everyday users with always-on AI agents. But does getting rid of local control reduce the power of the framework? Is KimiClaw actually useful for professionals, or is it a simplified version of a developer’s favorite tool?

Let’s discuss the architecture, possibilities and trade-offs.

# Understanding the OpenClaw architecture

To evaluate KimiClaw, we must first understand the engine it runs on. OpenClaw is not a language model. It is a gateway to orchestration – a platform that connects your preferred LLM to the operating system.

When you interact with a customary LLM, the architecture is completely reactive. You send a prompt, the model generates text and the interaction ends. OpenClaw changes this through four primary mechanisms:

// Be proactive by beating your heart

OpenClaw runs as a persistent background daemon with a configurable heart rate, typically waking up every 30 to 60 minutes. During each cycle, the agent reads the local file independently HEARTBEAT.md checklist, assesses whether background tasks require action, and executes them. It can crawl a competitor’s site, manage something like Gmail’s inbox routing system, or run a data pipeline while you sleep, only notifying you when a task is completed or requires human intervention.

// System-level execution

Since the framework resides on your computer, it has permissions to perform actual actions. It can run shell commands, control a web browser, read and write files, and manage Docker sandboxes. The text generated by LLM acts as a system control signal rather than a conversational response.

// Maintaining persistent Markdown memory

Classic web chats remove context when you close a tab. OpenClaw manages long-term state by constantly rewriting its own local configuration files. Basic instructions are written in a format SOUL.md file while the user’s facts and preferences are saved MEMORY.md. Before processing any up-to-date message, OpenClaw inserts these files into the context window, ensuring consistent recall of workflows and rules.

// Routing over ubiquitous channels

OpenClaw captures messages from apps you already exploit. Thanks to channel adapters, it normalizes input from WhatsApp, Telegram, Slack or Discord, funneling everything into a continuous session.

This architecture transforms AI from an oracle to a proactive background worker.

# Hardware bottleneck and Mac Mini performance

The power of local OpenClaw comes with real infrastructure requirements. In early 2026, the framework’s popularity sparked a notable launch on the Apple M4 Mac mini, which has become the de facto standard for running personal AI agents.

This hardware dependency occurred for several reasons. OpenClaw requires an always-on machine to maintain the heartbeat daemon and run cron jobs 24/7. Mac mini consumes minimal power when idle, making it a practical choice. Running an autonomous agent capable of executing terminal commands on a primary work laptop also creates security risks, including up-to-date threat vectors such as AIjacking. A dedicated headless machine allows users to safely isolate AI from personal data. macOS is also strictly required to route the agent via native Apple iMessage. Finally, Apple Silicon’s unified memory architecture makes it well-suited for running local models efficiently.

While this setup is effective, it requires purchasing dedicated hardware, managing Node.js environments, and resolving conflicts on the command line. For professionals who want automated workflows without becoming system administrators, this barrier is too high.

# Introducing KimiClaw: a cloud-hosted approach

This is the Moonshot AI friction point that KimiClaw targets. The platform allows users to run OpenClaw-style agents directly from a browser or mobile device, without the need for on-premises servers, intricate deployments, or the need for a VPS.

It takes the OpenClaw orchestration layer and moves it to a managed cloud infrastructure, transforming the platform from a standalone development tool to a software-as-a-service (SaaS) product. This is what it makes possible for data professionals and automation enthusiasts.

// Eliminate technical setup and guaranteed uptime

With self-hosted OpenClaw, your agent only runs as long as your computer is turned on and connected. Hardware failures, network outages, or simply closing your laptop kill your heartbeat. Because KimiClaw runs on Moonshot AI servers, your agent stays online constantly.

This reliability is most crucial for scheduled background tasks. If you designate an agent to run a data extraction script at five industry locations every morning at 4:00 a.m., KimiClaw will handle the execution without the need to maintain a physical server.

// Leveraging the Integrated Skills Marketplace (ClawHub)

To extend the capabilities of your local OpenClaw agent – for example, teach it to analyze analytical dashboards or execute Python code – you must manually install “Skills”. Managing them locally means dealing with dependency conflicts and version incompatibilities.

KimiClaw integrates with cloud-hosted ClawHub a marketplace home to thousands of community-created skills. When you assign a intricate task, KimiClaw can automatically select, install and connect the appropriate skills in the background. This allows the agent to combine web browsing, graph generation, and data analysis into a fully automated pipeline.

// Using built-in persistent memory and cloud storage

Local management of Markdown persistent storage files can become disorganized on multiple devices. KimiClaw provides a unified workspace with 40 GB of cloud storage. All files, PDFs, logs, datasets and reports generated by the agent are saved in one centralized hub. The platform supports the persistent long-term memory that made OpenClaw popular, so established rules, formatting preferences, and workflows are reliably implemented between sessions.

// Enable mobile and visual device controls

One of KimiClaw’s more noteworthy features is its mobile capabilities. Through the Android app, KimiClaw uses accessibility APIs to visually read your device’s screen. It can autonomously move between applications, touch, swipe and interact with interfaces just like a human would. This allows the agent to perform cross-app operations, reference data in unconnected mobile apps, and manage workflows natively on the phone – something local OpenClaw doesn’t offer out of the box.

# Weighing the trade-offs

KimiClaw is really useful for most users. Delivers the core value of an autonomous agent without the complexity of infrastructure. However, it is not a 1:1 replacement in every exploit case and it is worth considering the trade-offs honestly.

// Accepting local access restrictions

KimiClaw works like virtual hardware, providing an instant sandbox. You don’t have to worry about the AI ​​executing a destructive shell command on your local drive. But this security comes at a price. Because it is a cloud service, KimiClaw cannot control local applications or read files stored on your personal computer unless you actively upload them to your workspace.

// Considering data privacy

With OpenClaw self-configuration running on-premises, 100% of your data stays on your hardware. KimiClaw requires knowledge of agent memory, system messages, and generated data stored on Moonshot AI servers. For enterprise teams handling sensitive or proprietary data, this cloud dependency can be a factor in termination.

// Navigating differences in platform integration

While local OpenClaw on a Mac mini can route directly through the native Apple ecosystem, KimiClaw uses third-party messaging platforms like Telegram to communicate with an agent on a mobile device. For users deeply embedded in the Apple ecosystem, this is a significant gap.

# Verdict

OpenClaw has proven that giving AI quick performance and system-level access can transform the way personal productivity and data automation work. KimiClaw takes this framework and makes it available.

It’s a solid tool for professionals who need reliable 24/7 automation, web browsing capabilities, and persistent storage, but who don’t want to manage dedicated hardware or troubleshoot command-line interfaces. For engineers who need complete data sovereignty and local system control, self-hosted OpenClaw is still a better option. However, for practitioners who want to immediately deploy an automated background worker, KimiClaw will do the job at no additional cost.

Vinod Chugani is an artificial intelligence and data science educator who bridges the gap between emerging artificial intelligence technologies and practical applications for working professionals. His areas of interest include agentic artificial intelligence, machine learning applications, and workflow automation. Through his work as a technical mentor and instructor, Vinod has supported data professionals in skill development and career transitions. He brings analytical knowledge of quantitative finance to his hands-on approach to teaching. Its content emphasizes practical strategies and frameworks that professionals can implement immediately.

Firefighters in France are faced with a phenomenon never seen before in the region: clouds of fire. It’s a sign of the intensity of the flames and how climate change is making them more likely.

A wave of fires swept across France and Spain, forcing hundreds of thousands to flee. Before the firefighters on the front line stands the spokesman of the French firefighters’ association he told AFP is called “operational impossibility” – in other words, “a natural force beyond our control.”

As evidence of how out of control the fires in France have become, just look at the sky, where in some cases the fires have created their own weather. To create a fire cloud – which meteorologists call pyrocumulonimbus or the even more metallic-sounding cumulonimbus flammagenitus – you first need heat, of which fires have plenty.

But equally significant are dehydrated conditions near the ground and frigid, relatively humid conditions in the atmosphere. As the superheated smoke rises miles above the flames and into the frigid atmosphere, water vapor condenses around the ash particles, forming water droplets. (Yes, water vapor and water droplets are two different things.) As the air continues to rise, these water droplets eventually turn into ice crystals.

At this point, smoke and water droplets had formed a cloud, but unfortunately it was unlikely to provide any relief in the form of rain. Instead, these towering fire clouds can trigger lightning strikes that can ignite more fires and generate powerful downdrafts that reach the ground and further fan the flames. In the worst case, pyrocumulonimbus clouds can even produce tornadoes.

Pyrocumulonimbus clouds have been documented in the US, Canada and Australia, as well as several other places. However, there have been no documented fire clouds in France so far. While the country is no stranger to wildfires, the size and ferocity of this summer’s fires has been far from the norm.

According to data from the European Forest Fire Information System, last week was the most devastating week for forest fires in France in the last 20 years. This is more than double the previous record for this period, which coincides with precise satellite data. This isn’t an isolated bad week either. So far, more than 220,000 acres have burned across the country, six times the annual average. That’s 61,000 acres more than the previous annual record.

If this sounds like a familiar trend, unfortunately it is. The burning of fossil fuels has warmed the planet, causing more constant and destructive fires around the world. AND 2024 study found that the incidence of extreme wildfires more than doubled worldwide between 2003 and 2023. Six of the seven most extreme years have occurred since 2016.

After a slight weekend lull as temperatures dropped, temperatures are expected to rise again above 104 degrees Fahrenheit (40 degrees Celsius) in France and 108 degrees Fahrenheit in Spain by the end of the week. This means firefighters will have to deal with more extreme fire weather, including the possibility of cloud formation from the fires themselves.

The world could face the strongest El Niño since records began, increasing the likelihood of extreme weather events and significant long-term economic impacts around the world.

Novel analysis by Berkeley Earth climate scientist Zeke Hausfather paints a disturbing picture of a climate phenomenon unfolding in the Pacific Ocean. Models have consistently predicted that this year’s El Niño – characterized by hotter-than-usual waters in the eastern tropical Pacific – could be one of the strongest on record. Hausfather analysis the utilize of an El Niño forecast based on 14 climate models through July indicates that “this year’s El Niño is not only very likely to be the strongest event since reliable records began, but may turn out to be the strongest by a truly stunning margin.”

Climate scientists measure the strength of El Niño by observing temperatures in the Pacific region known as Niño 3.4, which covers the eastern and central ocean basin. The fresh analysis is based on 667 simulations and compares them to historical records of the most intense El Niño events since 1877. It also distinguishes El Niño intensity from background warming caused by fossil fuels, making it easier to make comparisons between super El Niños.

The model average shows temperatures in the Niño 3.4 region could peak at 3.6 degrees Celsius (6.5 degrees Fahrenheit) above normal. This number would exceed the previous record set during the 2015-2016 El Niño episode by about 0.8 degrees Celsius. It’s infrequent for records to fall by such a wide spread.

“For context, the gap between the strongest and the strongest fifth the strongest El Niño in the last 150 years has a temperature of only about 0.5°C. The models predict something beyond anything we have ever observed,” writes Hausfather.

Of course, what happens in the tropical Pacific does not stay in the tropical Pacific. El Niño changes the atmosphere, changing weather conditions around the world. The more intense the El Niño, the more intense its impacts (usually).

Record ocean heat in the Pacific, combined with warming from greenhouse gas emissions, could push global temperatures to a new record. However, the extra heat released by El Niño will take some time to filter through the vast climate system, which means next year we will see the full impact of this year’s Pacific heatwave on average global temperatures. According to his forecasts, the global temperature in 2027 may reach up to 1.7 degrees Celsius above the pre-industrial average, which is well above the 1.5 degrees Celsius threshold defined as a relatively safe level of global warming.

But the world won’t have to wait until 2027 to feel the heat of El Niño. Separate analysis by Carbon Brief shows that the chances are increasing for 2026 will set an annual heat record, with the probability increasing from 19 percent in April to 35 percent in July. Forecasters expect El Niño-related extreme weather conditions to affect various parts of the globe in the coming months. In fact, some effects are already noticeable: catches are down in eastern Pacific fisheries, and the Peruvian government imposed a ban on anchovy fishing this spring.

Hausfather’s analysis also revealed that this phenomenon is developing faster than usual. The comparison shows that this year’s El Niño is intensifying faster than the 1997–1998 event, another super El Niño. It’s also very different from the 2015 event, which began when the ocean was already showing signs of early warming – in other words, it got off to a brisk start. This is especially striking because this year started with La Niña, El Niño’s cooler-than-usual cousin, so it had some work to do before it even got started.

10 newsletters to keep you ahead in AI
Vinod Chugani · Artificial Intelligence · July 22, 2026
Curated newsletters offer the necessary framework for filtering the AI ​​noise, providing the sturdy signal, multi-faceted coverage covering daily news, in-depth technical research, political strategy and tool-building tools necessary to stay current in a rapidly evolving field.

Kaggle + free 5-day Agentic AI course by Google
Nahla Davies · Artificial Intelligence · July 22, 2026
The course is a follow-up to Google and Kaggle’s 2024 GenAI Intensive, which attracted over 140,000 developers and set a Guinness World Record for the largest virtual AI conference. The 2025 release narrows the focus from generative AI broadly to specific agents, which is a valid call considering how much confusion still surrounds what an agent even is.

Evaluating hallucinations in a language model with GraphEval
Iván Palomares Carrascosa · Language models · July 24, 2026
GraphEval is an evaluation methodology that uses knowledge graphs and natural language inference to detect and locate the specific source of hallucinations in the output of a immense language model, offering explainable error diagnostics.

Introduction to loop engineering
Shittu Olumide · Artificial Intelligence · July 23, 2026
Loop engineering is the practice of designing autonomous cycles that allow AI agents to iterate, validate results, and adapt based on environmental feedback, shifting the focus from optimizing individual prompts to building strong, self-correcting execution systems.

7 Best Alternatives to Claude’s Code for CLI Agent Coding
Abid Ali Awan · Programming · July 23, 2026
There are seven open-source alternatives to Claude Code that offer increased control over agent-based coding workflows, providing superior wire harnesses, local model support, and malleable integration ecosystems.

Run the Mythos enhanced encoding model locally using llama.cpp and Pi files
Abid Ali Awan · Artificial Intelligence · July 21, 2026
Deploying inference models locally using llama.cpp enables you to create quick, agentic coding workflows that can autonomously perform elaborate software development tasks without relying on external APIs.

5 key concepts of agentic AI that every engineer needs to understand
Shittu Olumide · Artificial Intelligence · July 24, 2026
Successful agent-based AI requires mastery of five fundamental engineering disciplines—tool usage, memory management, planning, coordination, and evaluation—to move systems from demonstration to reliable production environments.

5 free courses to take you from beginner to AI practitioner
Vinod Chugani · Artificial Intelligence · July 21, 2026
The most effective path to becoming an AI practitioner involves a sequential curriculum that builds expertise from basic mathematical logic and classical algorithms to understanding newfangled LLM architectures from first principles using open source tools.

A beginner’s guide to configuring Claude’s code for high-performance agent programming
Shittu Olumide · Programming · July 20, 2026
Tough, effective agent programming with Claude Code requires mastering a project-specific configuration hierarchy and leveraging advanced hooks to enforce consistent security policies and workflow automation beyond the core installation.

First steps with OmniVoice-Studio
Shittu Olumide · Artificial Intelligence · July 23, 2026
OmniVoice Studio is an open-source local desktop application that enables voice cloning and copying by executing all AI pipelines directly on the user’s hardware, ensuring data privacy without the need for external API keys or cloud services.

The idea that cancer may behave like an infectious disease, it may seem absurd; in humans, cancer cannot spread from one person to another. However, nature is still full of surprises. In exceptional cases, the cancer cells themselves can pass from one individual to another and continue to multiply in the up-to-date host. This uncommon phenomenon has been detected in animals such as dogs, Tasmanian devils and some species of molluscs, and it seems we can add one more to this exclusive club: catfish.

An international team of scientists has identified an infectious melanoma in brown catfish (Better shadowy) — the first documented case in fish and the first recorded in a freshwater ecosystem. Their research was published this month in the journal Nature.

An impossible pattern

The story began in 2012, when fishermen and biologists detected unusually huge numbers of catfish with black markings in Lake Memphremagog, a body of water that straddles Vermont and Quebec. Previous research has shown that these spots are malignant melanomas and that 23 to 37 percent of samples tested had these tumors, an unusually high rate for this species.

“It was surprising” he said Julie Dragon, a researcher at the University of Vermont and co-author of the study. “We wanted to know how bottom-dwelling fish develop cancer, which we associate with exposure to too much sunlight.”

Initially, it was thought that some toxic substance or pathogen related to the contamination might be the cause. One suspicion was that flooding caused by Tropical Storm Irene in 2011 degraded water quality by washing environmental pollutants into the lake. There were also concerns about the health of the lake, which provides drinking water to more than 175,000 people. However, initial genetic analyzes began to point in an unexpected direction.

It was then that researchers wondered whether the tumors originated from a single primary tumor that had learned to spread from one fish to another, rather than arising independently in each fish.

To test this unusual hypothesis, the team sequenced the complete genomes of tumors and fit tissue from diseased fish and compared the data with data from fit specimens from different populations.

The results showed that the cancer cells in the different fish were much more closely related to each other than to the animals in which they developed. In other words, the tumors shared a common genetic identity that was distinct from that of the hosts. Moreover, hundreds of thousands of genetic variants repeatedly appeared in the tumors, but were not present in fit tissues of the same fish.

Such a pattern would be virtually impossible if each melanoma arose independently. For comparison, the researchers analyzed hundreds of human melanomas and found that almost all of their mutations were unique to each patient. However, most of the mutations found in these fish cancers occurred in multiple individuals, which is a hallmark of infectious clonal cancer.

# Entry

The AI ​​space is moving so swift that the customary news cycle simply cannot keep up. By the time a major publisher publishes a paper on a recent generative model, the open source community has already reverse-engineered it, optimized it, and integrated it into dozens of recent applications. For data scientists, machine learning engineers, and technology professionals in 2026, the inbox has become the most valuable tool for staying current.

However, not all newsletters are created equal. The space is currently flooded with general AI-generated roundups. To actually get ahead of the competition, you need curated signals from real practitioners.

In this article, we’ll discuss the 10 best AI newsletters, ranked by the specific value they add to your workflow: Daily scans, research and technical dives, policy and strategy analysts, AND Builder ecosystem. We’ve also identified a core topic for each one, so you can subscribe to exactly the mix of news, code, and strategy you need.

# Daily scans

If you only have five minutes over your morning coffee to find out what’s been sent in the last 24 hours, these daily newsletters are your best bet. Each approaches the same rapidly changing data source from a different perspective: one optimizes for breadth, one optimizes for raw technical links, and the third optimizes for direct practical application. Together they cover the full spectrum of what it means to stay informed.

// 1. Artificial intelligence destroyed

Artificial intelligence destroyed is widely considered to be the world’s largest daily AI newsletter, with over two million subscribers. Founded by Rowan Cheung, it was built entirely with speed and scanning capabilities in mind.

• Why you should read it: This is the most powerful news journal available. Each issue presents the most significant model launches, product launches and industry developments of the day in a fast-paced, conversational format.
• Best for: Anyone who wants a complete picture of the AI ​​space in one go, without getting drowned in technical jargon – operators, founders, and those interested in AI alike.
• To combine: Artificial intelligence destroyed

// 2. TLDR IT

Part of a wider one TLDR newsletter family, TLDR AI is one of the densest and least promotional daily scans on the Internet. It is notoriously ruthless when it comes to formatting: all you need is a headline, a two-sentence summary, and a permalink.

• Why you should read it: Heavily skews technical. While other newsletters discuss boardroom dramas, TLDR AI links directly to recent GitHub repositories, ArXiv articles, and engineering blog posts.
• Best for: Developers and machine learning engineers who want raw links and technical signals, not long-winded narrative.
• To combine: TLDR AI

// 3. Superhuman artificial intelligence

Superhuman artificial intelligence complements the everyday ecosystem by focusing solely on application and productivity. It gained a lot of followers by answering one question: How can I actually apply this recent AI tool to get my job done faster?

• Why you should read it: Instead of focusing on model architecture or training flows, it provides daily tutorials, quick engineering tips, and workflow automation guides.
• Best for: Productivity enthusiasts, marketers, and non-technical professionals who want to apply AI as a practical tool today.
• To combine: Superhuman artificial intelligence

# Research and technical diving

If you want to understand the math, architecture, and changes that occur at the frontier model level, these weekly reads are a must-read. This section covers the full range of technical topics: accessible research frameworks led by a respected educator, practitioner-level open source model analysis, and state-of-the-art post-training research, including reinforcement learning from human feedback (RLHF) and direct preference optimization (DPO).

// 4. Party

Posted by Andrew Ng’s DeepLearning.AIThe Batch is a weekly benchmark report on artificial intelligence research frameworks. It is written with extraordinary pedagogical care, making convoluted academic breakthroughs easily accessible without sacrificing accuracy.

• Why you should read it: Contains selected research summaries from “Letters from Andrew Ng,” one of the most cited and repeated columns in AI media. Provides a physical, educational corrective to industry hype cycles.
• Best for: Students, practitioners, and data scientists who want research explained by authoritative educators, not one-size-fits-all journalists.
• To combine: Party

// 5. Before artificial intelligence

Sebastian Raschka is a respected machine learning researcher and author of several widely read machine learning textbooks. His newsletter, Before artificial intelligenceis a deep, technical dive into open-source Enormous Language Models (LLM), tuning techniques, and model evaluation.

• Why you should read it: Raschka actually tests the code he writes about. It covers competent parameter tuning (PEFT), low rank adaptation (LoRA), and optimization strategies with the rigor of a textbook but with the cadence of a blog post.
• Best for: Machine learning engineers who actively train, tune, and deploy their own open source models.
• To combine: Before artificial intelligence

// 6. Connects

As the industry has turned its attention to post-training, the issue of refining models after pre-training has become one of the most technically significant in the field. Connects has become a leading source of credible, stringent analysis on exactly this topic.

• Why you should read it: Written by Nathan Lambert, an AI researcher with deep experience in RLHF, it offers unparalleled insight into the open scale ecosystem and model evaluation metrics. Lambert writes with the authority of someone who performed these experiments, not just summarized them.
• Best for: AI researchers and engineers who want to deeply understand post-training pipelines and the open source model ecosystem.
• To combine: Connects

# Policy and strategy analysts

Artificial intelligence is no longer just a technological problem – it is a geopolitical problem. The bulletins in this section connect the dots between raw computing power and long-term global strategy. If your work involves regulation, national AI policy, or large-scale enterprise deployment, these two are indispensable reads.

// 7. Import AI

Written by Anthropic co-founder Jack Clark since 2016, Import AI is one of the longest-running and most prestigious bulletins in this field. It’s your best resource for understanding where AI research meets global policy.

• Why you should read it: Each weekly issue contains summaries of academic articles with original analysis of computing trends, national AI and governance strategies. Clark is notable for closing each issue with an excerpt from a brief AI-themed novel that has gained a following over the years.
• Best for: Researchers, policy professionals, and anyone following the strategic, long-term implications of the development of artificial general intelligence (AGI).
• To combine: Import AI

// 8. Median

Median stands out because it connects AI news directly to skills development. Published by learning platform DataCamp, it combines the week’s most significant data and AI developments with practical context and links to tutorials, courses and practical resources.

• Why you should read it: Instead of leaving you with information you can’t act on, it tells you what changed this week and what you should learn as a result. This wording makes it really useful for professionals trying to fill specific skill gaps.
• Best for: Data professionals and developers who want to systematically develop their AI and data literacy as the industry evolves.
• To combine: Median

# Builder ecosystem

For independent hackers, startup founders, and software engineers building the application layer of the AI ​​economy, these newsletters serve as their default social channels. They cover product AI with a speed and detail that no general-purpose publication can match.

// 9. Ben’s bites

If you want to know what AI startups are launching this week, read on Ben’s bites. It acts as the central nervous system of AI creators and the venture capital community.

• Why you should read it: Provides a quick look at recent AI startups, product demos, and niche tools created by independent developers before they hit the mainstream press.
• Best for: AI founders, product managers, and independent developers looking for inspiration on product and ecosystem trends.
• To combine: Ben’s bites

// 10. Hidden space

Hidden space is a groundbreaking publication in the field of AI engineering. Written by Swyx, it bridges the gap between customary software engineering and machine learning research in a way no other newsletter does.

• Why you should read it: Featuring highly technical essays and an accompanying podcast that interviews engineers building tools like LangChain, LlamaIndex, and state-of-the-art vector databases. The text assumes you can read code, which means the analysis goes several layers deeper than most industry publications.
• Best for: AI-first software engineers with a focus on API integration, pull-assisted generation (RAG), and multi-agent architectures.
• To combine: Hidden space

# Summary

Taking care of your inbox is one of the most effective ways to filter out the noise associated with the Generative AI hype cycle. The above ten newsletters cover the full range: breaking daily news, in-depth technical research, geopolitical strategy and the product development community.

You don’t need all ten. Start with one from each category, spend a month with them, and see which ones you actually open every time they land. These are the ones worth keeping. The rest can wait until you’re ready for more depth in a specific area.

It’s challenging to find a good signal. These ten are a reliable starting point.

Vinod Chugani is an artificial intelligence and data science educator who bridges the gap between emerging artificial intelligence technologies and practical applications for working professionals. His areas of interest include agentic artificial intelligence, machine learning applications, and workflow automation. Through his work as a technical mentor and instructor, Vinod has supported data professionals in skill development and career transitions. He brings analytical knowledge of quantitative finance to his hands-on teaching approach. Its content emphasizes practical strategies and frameworks that professionals can implement immediately.

Electrolyte powders, soluble tablets and fashionable ready-to-drink waters have moved from the pockets of athletes to one of the most popular wellness categories.

And it’s a crowded market, worth approx it is estimated that in 2025 it will amount to USD 39 billion worldwide and is projected to double over the next decade. Much of the growth was driven by wellness trends and influencer hype cycles such as Prime Hydration, created by YouTubers Logan Paul and KSI, which the company said exceeded $1 billion in sales within one year. There are dozens in other internet corners fame AND adjacent to the podcaster All brands promise ways to optimize your hydration experience.

These products are marketed as everyday essentials for everyone from office workers to gym goers and promise to combat brain fog, improve skin condition, eliminate fatigue and aid recovery. Here’s everything you need to know about your electrolyte obsession and when it’s really worth buying these products.

What are electrolytes?

Electrolytes are necessary minerals including sodium, potassium, calcium, magnesium and chloride that carry an electrical charge when dissolved in a fluid. This charge allows the body to do most of its basic work, such as helping muscles and nerves function properly. They work to ensure fluid balance inside and outside each cell, because a grave imbalance can cause headaches, muscle cramps, dizziness, fatigue and weakness. In more grave cases, severe electrolyte imbalances can interfere with heart function and require treatment.

Who needs electrolytes?

In most fit adults, the body regulates electrolyte levels perfectly well long before trendy flavored moisturizer sticks appeared on supermarket shelves. In fact, if you eat a sensibly balanced diet and drink water when you feel thirsty, there’s a good chance you’re already getting everything you need.

The pouch deserves its place in approximately three situations. Research from American College of Sports Medicine suggest that water alone is sufficient for exercise lasting less than an hour, but electrolytes become increasingly useful during longer sessions of sustained sweat-intensive exercise, especially in high temperatures and humidity. They are also useful for diseases involving high fluid loss, which is why oral rehydration salts are an established treatment rather than a wellness trend. You should also consider them during long periods of fasting when food that usually provides these minerals is not available.

Are there natural electrolytes?

There are many sources of electrolytes in everyday food. If you eat an appropriately varied diet (fruits, vegetables, lean proteins, whole grains, nuts, seeds), your electrolyte stores are constantly replenished without the need to consume specialized drinks or powders.

Tests by Beverage Hydration Index shows that milk-based drinks can keep you hydrated longer than plain water. The combination of protein, natural electrolytes and diminutive amounts of fat slows the rate at which fluid leaves the stomach, helping the body stay hydrated for longer. Coconut water can also replenish potassium.

Source of nutrition of Harvard’s School of Public Health notes that basically “any drink containing electrolyte minerals can be labeled as an electrolyte drink. Some are very expensive but contain only a small amount of these nutrients, while others may contain a high amount of just one nutrient, such as potassium or sodium.”

Can you have too many electrolytes?

For fit adults, an occasional electrolyte drink is unlikely to cause harm. However, daily consumption can sometimes be counterproductive.

Many electrolyte-containing foods contain significant amounts of sodium, even though most of the population already consumes much more salt than recommended. The World Health Organization estimates that global average daily sodium intake is more than twice the recommended limit, and approximately 1.89 million deaths per year can be linked to excess intake. Adding extra sodium on a regular basis may not be ideal for people dealing with high blood pressure or cardiovascular risk. Some preparations also contain significant amounts of added sugar, making them nutritionally closer to tender drinks than health products.

Anyone recovering from an illness or sweating during an August shift outdoors has legitimate reasons to supplement. For everyone else, the best approach to hydration remains unremarkable. Electrolytes themselves are non-negotiable. Electrolyte products are mostly a lifestyle purchase.

This story originally appeared on WIRED Middle East.