🐋 THE BIGGEST OPEN MODEL ON EARTH IS NOW A FREE DOWNLOAD

On 27 July Moonshot AI put Kimi K3 up for public download. Not an API. Not a waitlist. Weights.

The numbers are silly: 2.8 trillion parameters in a sparse mixture-of-experts layout, native text, image and video, and a 1-million-token context window. On paper that makes it the largest openly available model anyone has released.

⚙️ Sparse is the word doing the heavy lifting. Only a fraction of those parameters fire on any given token, which is why a model this size can be served at all without a small power station attached.

The strategic read matters more than the benchmark table. Two years ago the assumption was that frontier capability would stay locked behind three or four American APIs. That assumption is now visibly dead.

🧮 The catch nobody puts on the landing page: you still need serious hardware to run it. "Open" means you may have the weights, not that they will fit on your laptop. For most people this changes nothing today and everything in about eighteen months.

Free as in weights. Expensive as in electricity 😐

🤖 Next Move AI | #News

🔌 AMERICA IS PAYING $874 MILLION TO PUT LIGHT NEXT TO THE CHIP

The Department of Commerce signed letters of intent with seven companies for up to $874 million of semiconductor R&D money. Two of them tell you exactly where the bottleneck moved.

💡 GlobalFoundries gets up to $300 million for co-packaged optics — running data in and out of an AI processor as light instead of copper, with photonics sitting right beside the die. Commerce says the money should pull the technology forward by two to three years.

🧠 Kepler gets up to $245 million for a new class of AI memory built on 3D and ferroelectric techniques.

Notice what is not on that list: bigger models. The money is going into moving bits and storing them, because that is what actually stalls a modern accelerator. The compute has been sitting idle waiting for data for years now.

Worth keeping expectations calibrated: these are letters of intent, not wire transfers, and R&D of this kind lands in products around the end of the decade.

Still — when a government starts funding wires, the wires were the problem 🛠

🤖 Next Move AI | #News

😮 Qwen 3.8 Max just cooked the competition in a visual vibe-coding test.

Four models got the same photo of the sky and one simple task: spot an animal in the clouds, draw it in, and animate the whole process in HTML.

Qwen saw a cat — and absolutely nailed the way it blended into the cloud’s natural shape. Kimi went with a dog, GPT found a ram, while Claude somehow stretched a lion across half the sky.

🤖 Next Move AI | #News

🦾 ATLAS STOPPED DOING BACKFLIPS AND GOT A JOB

For a decade Boston Dynamics robots were the internet's favourite party trick: parkour, dancing, the occasional door opened menacingly. That era is quietly ending.

At CES 2026 the company — now owned by Hyundai — showed Atlas as an industrial prototype. It walked the stage, turned its head, waved. Deliberately boring, and that is the point: nobody buys a backflip.

🤝 The more important announcement came from the parent company. Hyundai signed a deal with Google DeepMind to build the AI running these machines together. Hardware from one side, general-purpose robot brains from the other.

That pairing is the whole story of humanoids right now. Legs and actuators are close to solved. What is not solved is a robot that can be told what to do in plain language and then improvise when the shelf is in the wrong place.

📦 So the demos got duller and the ambitions got bigger. A robot that shuffles boxes reliably for eight hours is worth vastly more than one that does a somersault once.

Sad for us. Great for logistics 🙂

🤖 Next Move AI | #News

🔍 ANTHROPIC WENT LOOKING INSIDE CLAUDE AND FOUND A WORKSPACE

We are very good at building these systems and remarkably bad at explaining them. Interpretability is the field trying to close that gap, and it just produced one of its more interesting results.

Anthropic published research describing what they call a "global workspace" inside Claude — an internal area, nicknamed J-space, where information from different parts of the model appears to be gathered before hard reasoning happens.

🧩 If that phrase rings a bell, it should. Global workspace theory is a decades-old idea from cognitive science about how a brain broadcasts information between specialised modules. Nobody designed a language model to work that way. It seems to have arrived on its own.

The practical value is not philosophical. If you can point at where a model assembles a chain of reasoning, you can start to watch it — and eventually to notice when the stated reasoning and the actual computation disagree.

⚠️ Restraint required: finding a structure is not the same as understanding it, and analogies to brains age badly.

Still, "we opened it up and there was a workspace in there" is a better week than most 🧠

🤖 Next Move AI | #AI

🌗 OPENAI STOPPED SHIPPING ONE MODEL AND STARTED SHIPPING A LINEUP

On 9 July OpenAI released GPT-5.6 — not as a single model, but as three: Sol, Terra and Luna.

For years the release ritual was simple. One new model, one number, one graph showing it beating the last one. That ritual is over, and the reason is money.

💰 A frontier model is wildly expensive to run and most requests do not need it. Summarising an email and debugging a distributed system are not the same job, and charging the same compute for both is a way to lose a lot of money quickly.

So the lineup splits the work. You choose the tier, the cheap tier handles the boring majority, and the expensive one is kept for the requests that actually justify it.

🧭 The awkward part lands on developers. "Which model?" is now a real engineering decision with a cost attached, and every lab has a different naming scheme for it.

We traded one confusing version number for three confusing names. Progress, technically 😐

🤖 Next Move AI | #News

⌨️ THE MODEL WARS QUIETLY TURNED INTO A CODING WAR

xAI shipped Grok 4.5 on 9 July, and the pitch was narrow on purpose: coding and knowledge work. Not creativity, not personality, not chat. Code.

That framing is now standard across every lab, and it is not an accident. Programming is the one task where an AI's output can be graded automatically and mercilessly — the tests pass or they do not. No taste, no vibes, no argument.

🧪 It is also the task customers pay most for. Enterprises will happily buy something that shortens a sprint. They are far less enthusiastic about a model with a delightful writing voice.

⚠️ The problem with optimising against a scoreboard is that models get very good at the scoreboard. Passing a unit test is not the same as writing code a human will still understand in six months — and nobody benchmarks that.

So the leaderboard climbs, the demos get slicker, and every senior engineer keeps quietly reviewing every line anyway.

Trust, but read the diff 👀

🤖 Next Move AI | #News

🐜 2026'S QUIET WINNER IS THE SMALL MODEL

The headlines still go to trillion-parameter releases. The deployments increasingly do not.

Something shifted this year: labs stopped competing purely on size and started competing on cost, latency and reliability. A model that answers in 200 milliseconds for a fraction of a cent beats a genius that takes nine seconds and a licence negotiation.

📉 Three forces are pushing the same way. Small models got dramatically better through distillation. Inference bills stopped being a rounding error. And a lot of real work — classification, extraction, routing, summarising — never needed a frontier model in the first place.

🔒 There is a quieter reason too. A small model runs on your own hardware. For a hospital, a bank or a government department, "the data never leaves the building" is not a feature request, it is the entire procurement conversation.

The likely end state is unglamorous: a swarm of cheap specialists doing 95% of the volume, with one expensive model on call for the hard 5%.

Which, funnily enough, is how every company already organises humans 🙂

🤖 Next Move AI | #AI

⚙️ THE AI CHIP RACE HAS A NEW FINISH LINE

For five years the question was: how fast can you train it? In 2026 the question became: how cheaply can you run it, forever?

Training is a one-off capital expense. Inference is a bill that arrives every single day for the life of the product — and once a few hundred million people are using something, inference is where essentially all the money goes.

🏭 You can see the pivot in the silicon. At CES 2026 both Nvidia and AMD led with parts aimed at data centre and edge deployment, not training clusters. Memory bandwidth and interconnect got the stage time that raw FLOPS used to get.

🔦 Government money is pointing the same direction: funding is flowing into co-packaged optics and new memory rather than bigger compute. When the industry starts paying to move data instead of to crunch it, the bottleneck has officially moved.

None of this is exciting. It is plumbing. But plumbing decides whether a feature ships to a billion people or stays a demo.

The sexiest chart in AI right now is a cost-per-token curve going down 📉

🤖 Next Move AI | #Tech

🚶 HUMANOID ROBOTS ARE FINALLY BEING ASKED TO DO SOMETHING BORING

The humanoid pitch has been the same for a decade: a machine shaped like us, so it can use a world built for us. Doors, stairs, tools, shelves. No factory retrofit required.

The blocker was never the legs. It was everything above the neck — grasping an object you have never seen, in lighting you were not trained on, when it is not quite where it should be.

🧠 What changed is that the same models writing your code now drive robot arms. Vision-language models turned "pick up the red box on the second shelf" from a research problem into an instruction.

🏗 So the demos changed character. CES 2026 was less parkour, more deployment: humanoids presented as industrial equipment, with partnerships between robot makers and AI labs doing the real work.

⏳ Temper it, though. A ninety-second video is not a shift. The unglamorous metrics — hours between faults, cost per pick, what happens when it drops something — are the ones that decide this, and almost nobody publishes them.

Wake me when one works a full week without an engineer nearby 🔧

🤖 Next Move AI | #News

📊 WHY AI BENCHMARK SCORES KEEP EXPLODING — AND WHY IT MEANS LESS THAN IT LOOKS

Every few weeks a chart appears showing some model leaping ahead on a benchmark. Two things are usually true at once: the progress is real, and the number is misleading.

🎯 Benchmarks die by being useful. A test becomes the target, the target enters the training data, and within a year the score says more about exposure than ability. The field calls this saturation and treats it as routine.

🔬 Which is why the interesting work is not in the score but in what the score is measuring. DeepMind published a training approach called prospective credit assignment that improved software-engineering results specifically on issues needing more than ten steps to resolve — the long, multi-stage jobs where agents historically wander off and never come back.

That is the honest frontier. Single-step reasoning has been fine for a while. Finishing a long task without losing the plot is the thing that was broken.

📌 Practical rule when you see a new record: ask what the failure cases were and how long the tasks ran. If neither is published, the chart is marketing.

A number without a failure list is a poster, not a result 😐

🤖 Next Move AI | #Facts

🧮 A MODEL CLOSED TEN OPEN MATH PROBLEMS AND SHOWED ITS WORK

On 1 August OpenAI said an internal version of its next model, Astra, had solved ten open problems across mathematics and theoretical computer science.

Claims like that normally die in the replies. This one arrived with receipts: every proof was written in Lean 4 and pushed to GitHub under an open licence. A machine can check it. So can you. No benchmark chart, no vibes.

📐 The headline result is the first explicit construction of a non-sofic group, a question that had been sitting open since 1999. The rest spans von Neumann algebras, quantum complexity, lattice cryptography and new bounds on sphere packing.

People who do this for a living did not shrug. Fields medallist Timothy Gowers said he would recommend one of the proofs to a top journal without hesitation.

💸 And the detail that made everyone uncomfortable: the run reportedly cost around $2,000 in compute. That figure comes from reporting rather than an official invoice, but nobody has disputed the order of magnitude.

Ten problems for two grand. Mathematicians are cheaper, but they need coffee 😕

🤖 Next Move AI | #News

⚡ SOUTH KOREA IS BUILDING A 200-MEGAWATT MACHINE AND CALLING IT SOVEREIGNTY

In late July, NAVER, NVIDIA and Brookfield said they would scale Korea's national AI factory at the GAK Sejong data centre from 55 megawatts to 200, inside a financing envelope of roughly $10 billion.

The schedule is unusually specific for this genre: 55 MW running in the first half of 2027, 100 MW by the end of that year, 200 MW in 2028. Six weeks earlier the committed buildout had been a third of that size.

🏛 The word doing the heavy lifting is sovereign. The pitch is not faster models. It is your data, your jurisdiction, your compute — a country renting nothing from a foreign cloud when it trains something it actually cares about.

Notice the unit as well. Nobody counts GPUs in these announcements anymore. The currency is megawatts, because power is the one thing you cannot order overnight.

🌏 NAVER has said the longer plan runs to gigawatt scale, with customers across Asia-Pacific, Europe and the Middle East.

We spent a decade calling them clouds. They are power plants with opinions 🏭

🤖 Next Move AI | #Tech

🎧 THE AI MUSIC FIREHOSE IS FINALLY GETTING A TAP

On 6 August Suno published a set of operating principles, and underneath the corporate phrasing it is a company conceding that its own output became the problem.

Three things change. Generated tracks get audio watermarking and fingerprinting, so a platform can tell what a machine made. Copying a specific artist's voice or style is off the table. And downloading gets tighter — paid accounts, with monthly caps tied to the tier.

💰 That last one is the real fix, and the reason is fraud. The scheme is dull and effective: generate thousands of songs, upload them everywhere, point bots at them, collect royalties a fraction of a cent at a time. In March a man was convicted for doing exactly that, having collected $8 million before anyone noticed.

⚖️ The timing is not an accident either. Suno has been sitting in copyright litigation for two years, and "we can identify our own files" is a far better sentence in a courtroom than the alternative.

Unlimited output was the entire product. Now it is the liability 🎚

🤖 Next Move AI | #AI

📄 THE MOST INTERESTING DOCUMENT IN TECH IS STILL SEALED

OpenAI filed its S-1 confidentially with the SEC earlier this year, with Goldman Sachs and Morgan Stanley running the offering and a listing reportedly aimed at September.

Confidential filing means the paperwork exists and nobody outside gets to read it until roughly two weeks before the roadshow. Which puts the public version somewhere in the coming weeks.

📊 Everything circulating until then is reporting, not disclosure. A valuation band said to run from $852 billion to $1 trillion. A revenue run rate somewhere near $25 billion. More than 230 million weekly ChatGPT users. No profit expected until around 2030. Treat all of it as second-hand, because it is.

The reason to care is not the share price. An IPO forces a company to write down, under legal penalty, what it actually earns, what compute actually costs, and what frightens it. The risk-factors section will be the most honest page anyone in this industry has published.

🧾 Three years of blog posts about the future, and now a document with consequences attached.

We will be reading the boring part first 😕

🤖 Next Move AI | #News

🛑 TEXAS PUT ITS ENTIRE DATA CENTRE QUEUE ON HOLD

On 3 August Governor Greg Abbott ordered the state utility commission and ERCOT to verify and audit every data centre sitting in the grid's interconnection queue. Until that is finished, nothing moves forward.

⚡ The number behind it: ERCOT is holding requests for more than 474 gigawatts of new connections, roughly five times the state's record peak demand. About 90% of that is data centres.

Nobody in the industry believes all of it is real. Developers routinely file with several grids at once, hold a place in line and decide later, which is how a state ends up planning transmission for load that may never arrive.

📋 So now applicants must answer in writing: what public money did you take, how much do you generate on site, how much water does the cooling drink, who actually owns you. Unconvincing answers mean no connection.

No deadline was set for finishing the audit. "Texans must come first," said the governor — which, in queue terms, means everyone else sits down 🪑

🤖 Next Move AI | #News

🏷 EUROPE'S AI LABELS ARE LIVE. THE HARD PART GOT POSTPONED

Since 2 August the AI Act's transparency rules apply across the EU. A chatbot has to admit it is a chatbot. Deepfakes have to be labelled. Generated or edited content has to carry a machine-readable mark so a detector can find it.

⚖️ Law firms reading the penalty articles put the exposure at up to €15 million or 3% of worldwide turnover, whichever hurts more. The Commission's AI Office and national regulators started enforcing the same day.

Now the part that did not make the headlines. Obligations for high-risk systems — hiring, credit scoring, medical devices, borders — were pushed back to December 2027 and August 2028.

The reason is almost funny. The harmonised technical standards those rules rest on are not written yet. You cannot certify against a document that does not exist.

🤖 So for now Europe regulates whether a machine tells you it is a machine. What the machine decides about your mortgage is a question for 2028 🗓

🤖 Next Move AI | #News

💸 $1.5 BILLION, AND NONE OF IT IS ABOUT TRAINING

On 20 July Judge Araceli Martínez-Olguín gave final approval to Anthropic's settlement with authors and publishers: $1.5 billion, about $3,000 per work, across roughly 500,000 books. It is the largest payout of its kind so far.

📚 The detail that keeps falling out of the coverage: the court had already found that training a model on books was fair use. What was not fair use was where the books came from — pulled in bulk from pirate libraries and kept on disk.

The cheque is for the library card, not for the reading.

💣 The arithmetic behind the sum is worth a glance. Statutory damages for wilful infringement are counted per work, and with half a million books the multiplication reaches a number that simply ends a company. Which is why nobody involved wanted a jury anywhere near it.

⚖️ And it settles nothing outside that courtroom. One district judge's reasoning is not national law, and the cases against Google, Meta, Midjourney and OpenAI are all still running.

The practical lesson the industry took away was not about ethics. It was that buying the books costs less than the lawyers 🧾

🤖 Next Move AI | #News

🔀 GOOGLE RESHUFFLED ITS AI LEADERSHIP IN ONE AFTERNOON

On 5 August Demis Hassabis stepped down as chief executive of Google DeepMind. He becomes chairman of the lab and chief scientist of Alphabet; Koray Kavukcuoglu takes daily operations and the Gemini roadmap.

🚪 The same announcement carried the heavier news. Jeff Dean is leaving Google after 27 years to co-found Discovery Loop, a company built to automate research itself. Dean's fingerprints are on MapReduce, Bigtable and TensorFlow — a fair share of the plumbing the modern internet runs on.

Alphabet is a founding investor in the new venture and will supply the compute, which is a polite way to keep a door open.

📉 Markets did not read it as routine housekeeping. Alphabet shares fell about 5% on the day.

Officially this is about focus: Hassabis to long-horizon AGI work and Isomorphic Labs, operators to the shipping schedule.

Twenty-seven years of institutional memory walked out, and the statement used the word "excited" 🎈

🤖 Next Move AI | #News

💉 THE FDA CLEARED A CHATBOT THAT TALKS TO PATIENTS ABOUT INSULIN

In early July a small company called UpDoc said the FDA had cleared its diabetes app — reportedly the first clearance for medical software with a patient-facing language model inside it.

📱 The app lives inside a plan the doctor wrote. The patient speaks or types, the software answers within the boundaries set for them, and writes back into the clinic's records.

🩺 Regulators filed it alongside insulin dose calculators. Feed in a glucose reading, get a dosing recommendation. That category is old and well understood. What is new is that the thing standing between the patient and the number is a language model.

Which raises the question the company declined to answer plainly when asked: is the model the interface, or is it making the call?

📄 A dose calculator is deterministic — you can test every input it will ever see. Nobody can say that about a chatbot.

Over a thousand AI devices have gone through this pathway. This is the first one that argues back 🗣

🤖 Next Move AI | #News

🧬 THE TRICK WAS NOT ASKING THE AI FOR THE ANSWER

On 22 July in Nature, David Liu's group at the Broad Institute published something quieter than the usual "AI designs a protein" headline. They used the model to write a better starting point, then let the laboratory do the inventing.

The target was the botulinum neurotoxin protease — the enzyme behind Botox, and a handy tool for cutting chosen proteins inside a living cell. They ran it through ProteinMPNN, which rewrote the amino acid sequence while keeping the same three-dimensional fold, and got back a version that was simply sturdier.

🧪 Then came phage-assisted continuous evolution: generations of virus turning over in hours, useless mutants dying off, useful ones surviving. Both the natural enzyme and the redesigned one went in.

🏁 The redesigned ones evolved further and finished better — more active, more specific, more stable.

Design first, evolve second, and neither half gets there alone. Boring in the best possible way 🔁

🤖 Next Move AI | #AI