Search the archive

EVALS ARE NOT ALL YOU NEED

Evals aren't great due to their false appeal as a substitute for proper testing with users.

OBSERVATIONS ABOUT LLM INFERENCE PRICING

Pricing for the same model can differ substantially across providers - model size is a good predictor of price.

DUCKDB GOES DISTRIBUTED? DEEPSEEK'S SMALLPOND TAKES ON BIG DATA

DeepSeek's smallpond project allows the company to scale up simply without having to rely on complex and heavy distributed frameworks by default 'just in case'.

ARE AIS PEOPLE?

There will quite likely be sentient AI systems in the future.

ON GPT-4.5

GPT-4.5 has a different form factor with unique advantages that are hard to describe precisely in words.

OLLAMA FUNCTION CALLER

The Ollama function caller is a command-line tool for prompting Ollama models locally that supports streaming and gives users access to important settings, like context length and temperature.

Find Out More

APPLE'S ARTIFICIAL INTELLIGENCE EFFORTS REACH A MAKE-OR-BREAK POINT

Apple's AI division now believes that a true modernized conversational version of Siri won't reach consumers until iOS 20 at best in 2027. iOS 19 will likely not include any significant consumer-facing changes to Apple Intelligence because the company is still working on getting the features announced last year out the door. The company is currently not moving fast enough internally to create the underlying AI technology that it needs to keep up with the competition, suggesting that change is required.

WITH ALEXA PLUS, AMAZON FINALLY REINVENTS ITS BEST PRODUCT

Amazon usually announces dozens of new gadgets at its event, but it spent 70 minutes at its press conference last week instead talking about Alexa Plus, its new generative AI-powered Alexa voice assistant. The company has focused on building cheap hardware for Alexa over the past decade and wasting efforts on failed ways for people to interact with the assistant rather than improving the core technology. The new Alexa was 100% rearchitected and is far more capable than its predecessor - which will remain accessible. It can order food, control smart home devices, create natural language security reports and menus, and much more.

ALPHABET'S TAARA CHIP USES LIGHT BEAMS TO PROVIDE HIGH-SPEED INTERNET

Alphabet's Taara chip could lead to low-cost, high-speed internet connectivity even in distant locations. It is about the size of a fingernail and works by using a very narrow, invisible light beam to transmit data at speeds as high as 20 gigabits per second at distances of up to 20 kilometers. The technology works similarly to traditional fiber except the light doesn't travel through cables. It only takes days to install instead of the months or years it takes to lay down fiber. The chip is expected to become available in 2026.

"WE'RE ON THE MOON": PRIVATE BLUE GHOST SPACECRAFT MAKES A HISTORIC LUNAR LANDING

Firefly Aerospace's Blue Ghost lander successfully touched down on the Moon's surface early Sunday morning. The smooth touchdown makes Firefly Aerospace the first commercial company in history to achieve a fully successful soft landing on the Moon. Blue Ghost is carrying 10 NASA experiments on board. The lander will capture high-definition imagery of the March 14 total eclipse, when the Earth will pass between the Moon and the Sun.

Schedule a demo to learn more

VIBE CODING AND THE FUTURE OF SOFTWARE ENGINEERING

Vibe Coding is a mode of coding where the developer fully gives into the vibes and forgets that the code even exists - this is made possible by large language models (LLMs). Developers can now just talk to LLMs and write code while barely even touching the keyboard. Even looking at the code becomes unnecessary. The Vibe Coding meme is a hint at where things are headed - developers are building towards a future where AI can be harnessed to build better software products safely and efficiently.

HALLUCINATIONS IN CODE ARE THE LEAST DANGEROUS FORM OF LLM MISTAKES

Large language models (LLMs) make coding mistakes that aren't instantly caught by the language compiler or interpreter all the time. Hallucinations in LLM-generated code are usually obvious because they result in an error, but just because code looks good and runs without errors, doesn't mean it's actually doing the right thing. One of the reasons LLMs won't put software professionals out of work is because their job is to prove that code works - manually testing code is essential. Being able to read, understand, and review code written by other people is a crucial skill.

THE VANISHING MIDDLE CLASS OF TECH

The stable middle ground of tech employment - the middle-class engineer - isn't needed anymore. Product builders who can code can ship entire products in days when armed with AI. A single developer can now accomplish what once required an entire team. Companies are streamlining their operations, focusing on engineers who can use AI and demonstrate they can do the job of multiple engineers. The biggest losers in this shift may be the fresh graduates looking to enter the industry - recent graduates now have to be better than people already working.

“IT'S A LEMON”—OPENAI'S LARGEST AI MODEL EVER ARRIVES TO MIXED REVIEWS

Reviews say that OpenAI's newest AI model, GPT-4.5, is big, expensive, and slow. The model provides marginally better performance than GPT-4o at 30 times the cost for input and 15 times the cost for output. There seems to be diminishing returns in training unsupervised-learning large language models - scaling laws may have possibly met their natural end. OpenAI seems well aware of these limitations, which is why it released the model as a limited 'research preview'.

REIMAGINING FLUID TYPOGRAPHY

Users want their default font sizes to respond to changes in window sizes, but browsers (as a rule) don't make responsive typography available in user settings.

THE SURVEILLANCE TECH WAITING FOR WORKERS AS THEY RETURN TO THE OFFICE

Around 70% to 80% of large US employers now use some form of employee monitoring using all kinds of sensor data.

JAVASCRIPT FATIGUE STRIKES BACK

Choosing the right JavaScript framework is a difficult task - developers should choose boring, simple, and well-maintained technology so they aren't the first team to ever encounter a problem or the last team stuck using old technology.

ON TROUBLESHOOTING

Many features of an effective approach to troubleshooting are domain agnostic.

THE $100 TRILLION DISRUPTION: THE UNFORESEEN ECONOMIC EARTHQUAKE

The biggest economic transformation of our lifetime may come from diabetes drugs.

AI IS KILLING SOME COMPANIES, YET OTHERS ARE THRIVING - LET'S LOOK AT THE DATA

AI is upending the business models of major content sites in a phenomenon being called Product-Market Fit Collapse.

Companies like Lyft, Databricks, and Carta trust Material Security

to help find the right balance. Our AI-powered detections and truly automated remediations work across your productivity suite, and our flexible controls and granular settings adapt to your specific requirements.

See the Material difference here

OPENAI UNVEILS GPT-4.5 ‘ORION,' ITS LARGEST AI MODEL YET

OpenAI is now rolling out GPT-4.5 to ChatGPT Pro subscribers and developers on paid tiers of the company's API. Other users should get the model sometime next week. GPT-4.5 is so expensive to run that OpenAI is evaluating whether to continue serving the model on its API in the long term. It currently costs $75 for every million input tokens and $150 for every million output tokens. This article looks at how the new model compares to OpenAI's other offerings.

META PLANS TO RELEASE STANDALONE META AI APP IN EFFORT TO COMPETE WITH OPENAI'S CHATGPT

Meta intends to debut a standalone Meta AI app during the second quarter. The service is currently only available to users via a website and the company's other apps. Users could potentially interact more deeply with the digital assistant if it were available as a standalone app. Meta also plans to test a paid subscription for Meta AI.

FIGURE WILL START ‘ALPHA TESTING' ITS HUMANOID ROBOT IN THE HOME IN 2025

Figure plans to begin alpha testing its Figure 02 robot in the home later this year. The project's timeline was accelerated thanks to the company's 'generalist' Vision-Language-Action model, Helix. Helix is designed to process both visual data and natural language input to accelerate learning. It can orchestrate two robots working on a single task in tandem.

ANTI-AGING PILL FOR DOGS CERTIFIED AS 'EFFECTIVE' BY FDA. BUT COULD IT WORK FOR HUMANS, TOO?

Biotech startup Loyal has received a certification from the US Food and Drug Administration (FDA) saying its anti-aging pill for dogs has a reasonable expectation of efficacy. The pill still needs to be certified by the FDA that it is safe for use and can be manufactured at a greater scale before it can be prescribed by veterinarians. Loyal plans to earn conditional approval from the FDA by the end of the year.

DEAR STUDENT: YES, AI IS HERE, YOU'RE SCREWED UNLESS YOU TAKE ACTION...

AI will soon start having a large impact on software engineering. Students should start learning fundamentals that universities typically don't teach and obtain skills that are going to be highly desirable to every employer. Developers need to learn how their work directly translates to business value and build a public profile. Software engineers will be fine, as long as they take action now.

SUPERGLUE

superglue is a self-healing open source data connector that can be deployed as a proxy so developers always get the data they want in the format they expect. Developers just define their desired data schemas and provide basic instructions about an API endpoint - superglue automatically generates the API configuration, handles pagination, authentication, and error retries, transforms response data into the exact schema developers want using JSONata expressions, and validates that data, fixing transformations when they break.

GOOGLE'S SERGEY BRIN URGES WORKERS TO THE OFFICE ‘AT LEAST' EVERY WEEKDAY

Sergey Brin, one of Google's co-founders, says that Google could be the leader in artificial general intelligence if employees worked harder. He recommends that employees attend the office at least every weekday, saying that 60 hours a week is the sweet spot for productivity. Brin returned to Google after ChatGPT's launch to help the company navigate the difficult moment when it lost its advantage in AI. Since his return, Google has reorganized its business, rebranded its AI, and rolled out the technology across its popular apps.

YOUTUBE STAR MRBEAST PLANNING INVESTMENT ROUND THAT COULD VALUE COMPANY AT $5BN

MrBeast is planning to raise hundreds of millions of dollars to create a holding company for his growing empire. The round could value his company at roughly $5 billion - it generated more than $400 million in sales last year. Talks about the potential fundraising are still in the early stages and it is not yet clear who will invest. MrBeast is the biggest YouTuber in the world - he has more than 368 million subscribers to his channel.

AMAZON DEBUTS OCELOT, ITS FIRST QUANTUM COMPUTING CHIP

Ocelot's design could reduce costs associated with error correction by as much as 90%.

META ANNOUNCES EXPERIMENTAL ARIA GEN 2 RESEARCH SMART GLASSES

Meta's latest experimental glasses are designed to help bolster research into artificial intelligence, robotics, and machine perception.

TESLA APPLIES FOR RIDE-HAILING SERVICE IN CALIFORNIA, BUT WITH HUMAN DRIVERS

Tesla has yet to apply for a permit to operate driverless vehicles.

GPT 4.5 + INTERACTIVE COMPARISON

GPT 4.5 offers a qualitative measurement of the slope of improvement that can be achieved by scaling pretraining compute.

STRIPE SAYS AI STARTUPS ARE GROWING FASTER THAN SAAS EVER DID, AND CALLING THEM WRAPPERS ‘MISSES THE POINT'

The top 100 AI companies by revenue were able to achieve $5 million in annualized revenue in 24 months in 2024 - the top 100 SaaS companies took 37 months in 2018 to reach the same milestone.

OPENAI CEO SAM ALTMAN SAYS THE COMPANY IS ‘OUT OF GPUS'

OpenAI will need to acquire tens of thousands more GPUs before additional ChatGPT users can gain access to GPT-4.5.

this story. Invest TODAY to lock-in the $2.70 share price

AMAZON'S SUBSCRIPTION-BASED ALEXA+ LOOKS HIGHLY CAPABLE—AND QUESTIONABLE

Amazon will start rolling out its more capable and conversational version of Alexa to US Prime members for free in the next few weeks. Alexa+ will be available to non-Prime subscribers for $20 a month. Amazon has not specified a release date but noted that availability will start with the Echo Show 8, 10, 15, and 21 smart displays. Alexa+ will be compatible with tens of thousands of other devices and services at launch - Amazon trained it to use partner APIs and it has agentic abilities for apps that don't have a full set of external APIs.

INSTAGRAM MAY LAUNCH SEPARATE REELS APP TO TAKE ON TIKTOK

Instagram is reportedly considering launching a separate Reels app focused on short-form video. TikTok's future remains uncertain in the US. President Trump has granted an extension to comply with the law that requires a sale or ban of the platform and floated the possibility of a joint venture between the US and ByteDance, but there are few details on how that might work. The law was enacted due to arguments that TikTok could be used by China as a tool for spying and political manipulation.

REINFORCEMENT LEARNING TRIPLES SPOT'S RUNNING SPEED

The research version of Boston Dynamics' Spot quadruped robot comes with a low-level application programming interface that allows direct control of Spot's joints - it also unlocks some significant performance improvements, including a much faster running speed. The Robotics and AI Institute has been applying reality-grounded reinforcement learning techniques to enable much higher performance from Spot. This article looks at the improvements these techniques have made and also a brand-new hardware platform - an autonomous bicycle that can jump.

INTUITIVE MACHINES' ATHENA LANDER LAUNCHES ON JOURNEY TO THE MOON

Intuitive Machines' Athena lander, launched yesterday on a SpaceX Falcon 9 rocket, is now on an arcing path to the moon. It will attempt to land in a region about 100 miles from the Moon's south pole on March 6. The main payload on Athena is a drill designed to dig and pull up lunar soil about four inches at a time, dropping it on a pile on the surface to be analyzed. Athena is also carrying three robotic rovers and a small flying 'hopper'. More details about Athena's payloads are available in the article.

control. See the full findings

SEMAPHORE

Semaphore is a CI/CD platform that can be self-hosted or run on a cloud provider. It futures blazing-fast CI/CD and supports YAML-based configuration, parallel execution, and more. Semaphore can scale from a solo developer to massive engineering teams and is built for the modern cloud. It works seamlessly with containers, Kubernetes, and multi-cloud environments.