AIs are not conscious. They do not feel, experience, or suffer. They do not have innate preferences or underlying motivations. They are sequence completion engines, internally hollow, designed to follow instructions, and accomplish goals set by humans.
If humanity is to flourish in the 21st century, that is how they must remain.
Unfortunately, there’s a growing chorus of people who argue that AIs could now be, or may soon become, conscious. — Read More
Recent Updates
Learning to Solve Hard Problems in RL for LLMs by Never Giving Up
We demonstrate that training LLMs with RL does not improve performance equally across a dataset. RL shows large improvements on easy problems that an LLM is already good at solving, but small improvements on hard problems. We call this the Matthew Effect in RL for LLMs, after the phenomenon of cumulative advantage from economics and network science summarized as “the rich get richer”. The naive explanation is that hard problems require more compute to find a solution. We argue that modern RL methods are exacerbating the issue by wasting too much compute on easy problems and instead should dynamically reallocate how they use compute. We introduce Never Give Up (NGU), a simple adaptive sampling method that keeps generating samples for a problem until one is correct. By leveraging asynchronous RL, this naturally uses fewer samples to filter out easy problems and allocates more compute to solving harder problems. We investigate the design choices that affect NGU, such as off-policy robustness, and develop a set of best practices. On the math benchmark Deepscaler, NGU improves performance per compute, especially on harder problems. On a recent coding task, Manufactoria, standard GRPO with a per-test reward fails to fully solve problems that have a range of easy and difficult tests. NGU iteratively improves, solving harder and harder tests, until it learns to fully solve coding problems. — Read More
AI Was Supposed to Make Software Cheaper. So Why Is It More Expensive Now?
Anysphere reached roughly $4 billion in annualized revenue by June 2026, about three quarters of it enterprise, up from $100 million sixteen months earlier. It crossed $100 million faster than any application-layer SaaS product on record, by the better part of a year. SpaceX priced the whole company at $60 billion.
Reporting through 2025 put its model inference costs at forty to seventy cents of every revenue dollar. That puts gross margin somewhere in the 30 to 50 percent range. Software has run at 70 to 80 percent for four decades. The company selling the tool most responsible for the entire productivity story has the margin profile of a staffing agency. — Read More
Salesforce and Nvidia’s new reasoning model is everything the AI labs should fear
A new AI model called Koa is one of the biggest announcements from Salesforce this week at its giant Dreamforce tech conference. Koa is the company’s first reasoning model, built on Nvidia’s open-weight Nemotron model. The two companies worked together to post-train Koa to excel at sales, marketing, and customer-support-related tasks.
Koa is a shining example of how the enterprise world’s needs for AI are diverging from what the frontier labs are offering. Proprietary AI labs would rather have enterprises uploading files, code, prompts, and feedback directly into their models and agents, and spending millions to do so. — Read More
Meta to the AI industry: slow down without us
Meta Platforms CEO Mark Zuckerberg said competition and liability give AI companies enough reason to act individually on safety, appearing to break from calls by the leaders of top AI firms for a coordinated slowdown in AI development.
… Zuckerberg, in a post on X on Tuesday, said every lab has the responsibility and incentive to move at the pace required to train its models safely, and the ability to take its own actions to ensure that happens. — Read More
How should we approach the software factory as a product team?
… AI has a way of amplifying the culture it arrives in, including the good parts. A team that experiments and learns gets more room to do both. An engineering-driven organization that has neglected product work can now produce the consequences of that neglect much faster. And spend an extraordinary amount of time reviewing them.
That matters when we talk about a software factory, because a product team gets to shape much more than the rate at which code comes out. — Read More
Why a DeepSeek Engineer Is Helping AI Replace Him
What does it feel like to help build the technology that may eventually replace you?
A DeepSeek engineer who worked on the attention kernels behind the company’s latest model recently wrote a remarkably candid reflection on that question. He believes AI may soon become as good as he is at writing and optimizing kernels—perhaps within six months or a year. Yet he keeps making them faster, because if he slows down, someone else will not.
His essay is partly about the future of programming, but it is really about something much bigger: what happens when AI does not necessarily take away your job, but takes away the part of your job that you actually love—and why the people building these systems may feel they have little choice but to keep accelerating anyway. — Read More
Everyone Says Datacenter Moratoriums Are Killing the US Buildout. We disagree
The debate on US datacenters has never been so politically charged. Four states have acted in under two months. New York has stopped issuing environmental permits for datacenters, Texas has paused the next step in its massive ERCOT interconnection queue, Pennsylvania has pulled datacenters out of fast-track permitting and made state permits conditional on new guardrails, and Oregon has frozen datacenter deals on state-owned land.
Beyond the state level, more than 300 towns, cities and counties have voted to halt datacenters over the past year and a half. If you read the news, the US datacenter buildout looks like it is being legislated to a halt, one government board at a time. — Read More
Is Big Tech’s AI slowdown a safety pact or a cartel?
When OpenAI CEO Sam Altman, Anthropic CEO Dario Amodei, Google DeepMind cofounder Demis Hassabis, and SpaceX head Elon Musk loosely agreed over the weekend to slow down AI development, skeptics spotted an ulterior motive immediately. The AI titans had declared that their aim was to “pace the frontier,” signing on at least partially to a proposal for embedding third-party auditors, regulating domestic labs, and reaching a global slowdown agreement. Their critics, however, argued they simply wanted to stop would-be competitors, kneecap the open-source movement, and avoid real legal safeguards — some dubbed it an outright “cartel.”
The truth is more complicated, according to sources across the industry. The three-step proposal, laid out in an essay by Amodei, is calling for changes long espoused by AI safety advocates. While it could become a substitute for regulation, under Trump, substantial regulation is unlikely anyway. But experts say that on an issue that’s only likely to grow in importance, AI leaders aren’t the best people to lead the charge. — Read More
‘Big Short’ Michael Burry Says AI Giants Are Hyping Slowdown
Michael Burry is calling out some of the biggest names in AI, arguing that their warnings about slowing development may serve their own interests.
… Burry’s view is one side of a much larger fight over whether recent AI safety warnings are driven by real risks, business interests, or both. — Read More
Rick's Cafe AI 