Welcome back! This week, we have a huge theme: open models, with several news items putting them at the center of the entire industry, including new open models achieving SOTA results, research that will make their adoption easier, and even some companies in the space publicly attacking Frontier Labs.

We’ll also cover new research, like Meta’s mind reader, market data, and new products and models, including an incredibly realistic AI-generated video and a chores robot that may come to you as soon as this year.

Enjoy!

I’ve long sustained that the future of enterprise AI is companies training open models in their own data while giving an enthusiastic ‘goodbye’ to OpenAI and Anthropic, at least for the vast majority of enterprise use cases.

And it seems several companies agree with me, including key players in the AI space like Palantir and Microsoft, and are becoming increasingly vocal about it (more on that below).

The appeal of lower costs, better governance, and tighter security makes this a no-brainer once open models are reaching a level of capability that warrants adoption.

The strongest argument against this idea has always been obsolescence, meaning why would I spend a couple thousand dollars fine-tuning an open model if it’s going to be obsolete by next week?

And while that argument is already not particularly strong when you realize that it’s okay to have legacy models for some tasks because these tasks do not need new levels of intelligence (e.g., I continue to use Gemini 3 Flash for a lot of my enterprise tasks despite this model being more obsolete than dinosaurs relative to the frontier, but it’s way cheaper), this new work by Ramp, an expense management company that is starting to look more and more like an AI Lab these days, proves how you can also ‘port’ your fine-tuning efforts from one model to the next.

But how? Leveraging one of the most fascinating pieces of research I’ve ever come across, Text-to-LoRA, which I’ve talked about in the past.

Fine-tuning is the practice of taking a pre-trained model and retraining it for a specific task you’re interested in. This is great because the model becomes incredibly good at that task (at the expense of losing performance in other areas).

The problem with fine-tuning is that you’re still retraining a large model, which can incur prohibitive costs, especially considering that the industry moves so fast that new, superior models come out every week, reducing the incentive even further because getting a return on that training run feels very complicated.

An alternative is to use LoRAs, low-rank adapters. My previous link goes into detail, but the idea is that most tasks that a model has to learn are “low rank,” meaning only a very small subset of the model’s global parameter count has to be modified to learn the task. Hence, the idea is to only train a small portion of the model.

So what is LoRA training? Simple: train a tiny external adapter and add it to the model whenever it’s working on that task. This is not only cheaper but also leaves the open model untouched, allowing a single model to work with potentially hundreds of adapters, depending on the task, an architecture used in compute/memory constrained environments (e.g., Apple Intelligence).

You can think of LoRA adapters as giving a plumber a wrench for that task and only for that task. While you need to teach the plumber to use it, you don’t need to rewire the entire plumber’s brain; it just requires a small, quick learning process for using the wrench.

Importantly, the wrench is only required for wrench stuff, so it doesn’t become a part of the plumber’s “being,” and they can simply drop it if it’s not required for the task.

LoRA is incredibly effective and a standard for training. However, it still requires a training run and can be expensive relative to the base model's time-to-obsolescence (i.e., your base model might become ‘dumb’ relative to what’s available in the market pretty quickly).

For this, Sakana AI had a great idea called text-to-LoRA, using an AI to generate the LoRA conditioned on text.

For example, you can describe the task “write emails using formal language”, and this model automatically generates a LoRA adapter that, when added to a model, makes it write emails always in formal language.

Incredible, right? No training run required, fast iteration, and it works. What’s not to like?

Well, one problem remained: Text-to-LoRA outputs are not model-agnostic because they are trained alongside the chosen base model. In simple terms, they learn to generate adapters for a particular model.

Here’s where Ramp’s PorTAL comes in. Cutting to the point, they’ve managed to create a model-agnostic Text-to-LoRA that can port adapters from one model to the next.

For example, say you’ve trained (or even automatically generated) an email-filtering LoRA for your Qwen3.6 35B model… and Qwen4 arrives the next week and is considerably smarter, to the point that it’s worth the switch.

Instead of ditching the previous base model and writing off the training, you use PorTAL to transfer the LoRA to that new model. There’s some training required, but it’s minimal as the largest portion of the PorTAL system is model-agnostic.

The results are very promising, showing that the system recovers 98% of the per-task LoRA performance on an unseen base model. I know, that’s a lot of jargon in one sentence.

In layman’s terms, they prove that the PorTAL system can take an ‘unseen’ model and generate a LoRA with minimal training and cost, matching the effort required to train a LoRA adapter for that model and task from scratch.

I have a strong belief that many in this industry would consider very bold: much of Frontier Lab's revenue can be explained by customer unsophistication.

Which is to say, most Anthropic/OpenAI enterprise customers use their models because they don’t yet know how to build robust open model pipelines and workflows.

The math is already there; open models give you greater control, better cost management, and a fully sovereign AI stack. The problem is that enterprise leaders aren’t yet aware of this and aren't operationally capable either.

But once frontier token prices force companies out of their bubble and into real AI engineering, I believe Anthropic and OpenAI will be in a world of pain.

Besides liquidity, it’s this sophistication curve that makes these labs so eager to go public as soon as possible. Once the word is out, investors will find it much harder to underwrite today’s valuations on the basis of distant, increasingly uncertain revenue growth.

The “own your AI stack” trend is gaining huge momentum at the worst time possible for these Labs.

After seeing this video, I finally accept that I can no longer distinguish AI videos from real videos.

It’s impossible, at least to me, as models like SeeDance 2.0 have crossed the uncanny valley (the feeling you get when you see something that looks almost human but not quite) into something that is simply indistinguishable.

The results are simply incredible. But I want to remark on something that is often ignored in these situations: the incredibly beautiful prompt, also linked, that led to that video. That is also art; this is also something complicated most people can’t achieve because most people can’t prompt this way.

What I’m implying is that there’s still something human in all of this, there’s a craft here, and that the capacity to imagine and describe what needs to be generated is a skill like any other. After reading the entire prompt, I can assure you I don’t have the vocabulary or the creativity to create something like that, at least not today.

My point is that AI does elevate what the average human can create. But it further elevates the experts in the given domain. People assume AI destroys leverage, but I’m beginning to suspect it will simply give experts more leverage than ever.

It’s also interesting to see how AI videos are much more indistinguishable from real videos than AI-generated writing is. One possible explanation is that AI-generated content is everywhere, so we have been exposed to it so often that we can recognize when AI is involved.

It could be that AI videos, once they go truly mainstream, will suffer from the same issues, as AIs fail to generate meaningfully different results beyond their average output. Who knows.

As reported by BleepingComputer and other sources, users say Anthropic’s relaunched Claude Fable 5 feels “nerfed” after its return with stricter safety controls. The complaint centers on coding and debugging tasks, where some users report more refusals, fallbacks, and weaker results.

We now even have benchmark confirmation, with percentage changes that are quite dramatic.

Anthropic says the main change is an added safety classifier, which is, of course, much more stringent than the one that the USG and Amazon jailbroke to meet the USG’s requisites.

When Fable 5 flags a request, it may be blocked or routed to Claude Opus 4.8 instead. Anthropic says this was added after concerns about cybersecurity misuse and blocks the reported bypass behavior in over 99% of cases.

Independent benchmark claims are mixed. ModemGuides reports that one BridgeBench rerun showed sharp drops after Fable 5’s return, including debugging falling from 86.2 to 25.9, but notes this may reflect classifier blocks or fallback behavior rather than weaker underlying models.

This is how companies lose the mandate of heaven. At some point, we’ll have to open the can of worms of classifiers being legal. How is me getting charged for something (Opus 4.8) I did not intend to purchase because I was looking to use Fable 5?

I understand this is obviously in the terms of service, but it’s an extremely arbitrary and unclear event that should raise questions about whether the user is being scammed. If enough users feel scammed, is that enough evidence of a scam? I have literally no idea; I’m just throwing this out because I, for one, simply reject using these systems if I can’t foresee what model I will be served.

A new, interesting piece of research by Ramp and Revelio Labs, echoed by the Financial Times, has found a striking correlation that discredits the idea that AI doesn’t cause massive layoffs.

If anything, correlations show the opposite.

According to their analysis, companies adopting AI have a 10.2% average increase in headcount two years in, versus those with low AI adoption metrics showing basically no growth.

But before the statisticians in the crowd defame me for confusing correlation with causation, I am not, and neither are these researchers.

Based on the information I’ve shared with you so far, you might interject that this does not prove causality and that other variables may be at play, explaining the difference.

For example, it’s reasonable to assume that high-adoption companies also tend to be enterprises with higher growth, better financing, larger size, and other factors that could explain increased headcount.

The researchers acknowledge this and apply a difference-in-differences method. In simple terms, those other variables we’re discussing are also commonly found in companies that were late adopters, so what they’ve done is compare to this group.

In other words, while comparing adopters vs non-adopters is a bad analysis to isolate the AI effect for the reasons stated above, it’s much more reasonable to compare early adopters to late adopters, because both share many of these confounding attributes, and the impact on headcount during the time difference can be used to isolate the effect of AI.

This way, while not categorically claiming causality, as that would require purely randomized trials (i.e., giving AI at random to some companies and not to others, and seeing if AI causes a difference), it’s a much more realistic comparison, essentially making this “AI effect” be measured as follows:

AI effect ≈ employment growth after AI adoption among adopters − employment growth over the same period among similar not-yet-adopters

And the results are pretty good, showing a consistent pattern: early adopters’ headcount growth outpaced late adopters' considerably during the period when the former were using AI while the latter weren’t.

The data also shows some counterintuitive results that may surprise you. For instance, it’s reasonable to assume that one reason one company may be hiring faster than the other is its size. Smaller companies, like start-ups, usually grow faster and are more eager to hire.

However, the average headcount of the low-adopter group is much higher than that of the “never-adopters,” and it shows higher headcount growth. Yes, high adopters are generally smaller, but arguing that this is all a ‘large vs small’ comparison makes no sense.

Meta has introduced Brain2Qwerty v2, a non-invasive AI system designed to decode typed sentences from brain activity.

The model was trained on roughly 22,000 sentences from nine volunteers, each recorded for about 10 hours while wearing a magnetoencephalography, or MEG, device and actively typing on a keyboard. Meta reports that the system reaches 61% word accuracy on average, rising to 78% for its best participant.

The key point is that Brain2Qwerty is not a general thought-to-text model. It does not read arbitrary thoughts and convert them into language. In the experiment, participants were shown sentences, briefly memorized them, and then typed them on a QWERTY keyboard while their brain activity was recorded. The system learned to reconstruct the sentence from the brain signals associated with that typing task.

Therefore, the training signal is obtained by comparing the system's predictions with the participant's actual typing. If the model assigns a low probability to the correct sequence of letters or words, the error is used to update the model.

Over many examples, it learns which patterns in the brain recording tend to correspond to particular typed characters, word boundaries, and sentence structures, making the system closer to “brain-assisted typing reconstruction” than pure “mind reading.”

This is similar to what Neuralink is doing, with the difference that Neuralink’s chips are invasive in order to capture a less noisy signal that doesn’t lose strength by crossing the skull and the scalp.

I’ve talked about Meta’s weird desire to read human thoughts until you realize it makes incredible sense to them.

Imagine they could decode what you want based on how you interact with their platforms, like a reverse engineering process to what we have described; if they can decode how your brain is activated while using Instagram, they can then predict your needs by knowing what you like, what you dislike, and everything in between; the ultimate ad-targeting platform.

Sounds great, right?… Right?

In the last few days, several prominent companies upstream and downstream of the closed AI Labs, Palantir, Microsoft, and TogetherAI, as well as rivals like Mistral through its CEO, have voiced very strong opinions, especially Palantir, in favor of open models, or, more clearly: companies should not outsource their operational learning loop to a frontier-lab token API.

Microsoft is doing so with its Frontier Company/Frontier Tuning push, which focuses on helping enterprises build customized AI systems around their own data, workflows, and business goals, often within the customer’s environment and with greater model flexibility. Reuters framed this as Microsoft helping companies move away from dependence on a single AI provider, such as OpenAI or Anthropic.

Palantir is partnering with NVIDIA to onboard its Nemotron models to its Ontology platform and is being less orthodox about its opinions on Anthropic and OpenAI.

Karp has criticized “tokenmaxxing” and warned that companies risk giving away their IP, alpha, and operational knowledge to external LLM providers. His position is that the enterprise’s data, ontology, permissions, and workflow logic should remain under the company’s control, with models treated as replaceable components.

And from a more official channel, the company itself published a statement in favor of AI sovereignty, clearly stating that controlling the models' weights means controlling your fate.

This is, of course, a very self-serving narrative for all these players mentioned above, but they nail it in their own way. They are correct, and you should be scared straight from trusting your IP to these Labs.

But all I can say is that I feel vindicated. It’s finally happening. The transition to open models in the enterprise looks unstoppable, something I’ve been calling for years, even before Thinking Machine Labs came out with their RLaaS service, by that time, I had already made up my mind this was the future.

Not because I’m a genius, but simply because I paid attention to AI’s history, and except for the few recent years, the decades-long story of AI has always been about open, deep models. In other words, AI research has always been open, and the outcomes have always been task-specific; it’s like we’re returning to 2016 AI.

Yes, foundation models like the ones we have today are good at various tasks, but great at none, the opposite of what enterprises need; they don’t care that their customer support agent is also an expert cake baker, but they need it to be the best customer support agent possible.

Besides, companies are realizing the extreme stupidity of outsourcing their entire AI stack to third-party companies that not only access their IP (Anthropic literally publishes research classifying how users use Claude) but actively build downstream competitors using their IP, as happened to Figma with Claude Design and to pharma companies with Anthropic’s new drug-discovery initiative.

In the case of Figma, Anthropic’s Chief Product Officer was a board member and resigned only three days before Claude Design launched. Do with this information what you wish, but I have a very clear idea of how I would feel if I were Figma.

And to cut Anthropic some slack, OpenAI and Google do the same thing.

Luckily, open models are finally good enough (see the first news in the product section below), so you can train fully sovereign solutions on your data and achieve state-of-the-art performance at more than 10 times (or even up to 50 times) lower cost.

As published by TechCrunch, Anthropic is reportedly discussing a custom AI chip with Samsung.

The talks are still early. According to TechCrunch, Anthropic has not yet decided exactly what the chip would be used for, how powerful it would be, or how it would fit into its servers.

Anthropic also emphasized that its compute strategy will continue to rely on a diversified hardware stack that includes Google, Amazon, and Nvidia.

Better late than never. One intriguing thing for me here with OpenAI/Anthropic hardware initiatives is that they make sense on a cost basis; progressively lowering costs/token, but they could hurt them badly in accounting terms.

Right now, they are mostly avoiding depreciation costs and CAPEX, and simply renting compute from their own investors (Hyperscalers), who are gladly offering low rental rates because they can recognize that usage as AI cloud revenue and RPO numbers (north stars for many Hyperscaler investors).

This is what has allowed Anthropic to claim “adjusted profitability” (excluding stock-based compensation) for this quarter. But this much-celebrated milestone hides a problem: Anthropic can only claim this because there’s a fool behind them paying the real bills and suffering the massive cash flows (i.e., Amazon and Google).

But the moment the GPUs are yours, your company is no longer margin-focused and suddenly much more free-cash-flow-focused because you have significant cash outflows.

If you judge Anthropic or OpenAI by free cash flow rather than margins, the picture changes completely: in AI, margins are acceptable, but cash flows are horrendous, so these Labs might end up looking worse to investors the more they move upstream into owning the infrastructure.

On June 29 and 30, 2026, more than 60 companies listed on China’s A-share markets released announcements on abnormal stock price fluctuations and risk warnings.

The majority of these companies operate in the semiconductor industry and its supply chain, including chip design, manufacturing equipment, and materials. The announcements were filed with the Shanghai and Shenzhen Stock Exchanges.

They were triggered by sharp increases in share prices in recent trading sessions. Many companies reported cumulative gains of 60 percent or more over periods such as the prior 20 trading days, with some noting even larger rises over 10 or 30 days.

These gains exceeded the performance of broader market indices, including the STAR Market Composite and STAR 50 indices. In the filings, rolling price-to-earnings ratios—the valuation multiple relative to earnings; the higher, the more highly valued a stock is—were frequently cited as significantly higher than industry averages for the computer, communications, and electronic equipment manufacturing sector.

Examples include GigaDevice (stock code 603986), which noted that memory chip prices were at historical highs and could experience a considerable pullback, and Jiangsu Aisen Semiconductor Materials, which reported a 61 percent gain over 20 trading days alongside a rolling P/E ratio of 233.86 times.

These are absolutely insane numbers. For reference, the companies in the S&P 500 have average trailing and forward P/Es of 20x and 32x, respectively. Even a company as richly valued as Palantir has a PE of 135, way smaller than some of the numbers we’re seeing in China.

But what on Earth is going on?

Chinese listed companies (especially in the semiconductor supply chain) are required to issue these when their stock prices move abnormally (large gains over a short period).

One or two issuing such risk disclosures doesn’t say much, but when 60 do so in the last two days, something’s off.

And when people looked more closely, they found that all these companies were related to the AI and semiconductor industries, highlighting incredible exuberance.

It’s becoming quite clear that the AI markets are overheating, and many investors will see this as an opportunity to sell. Will they be right? Who knows.

And talking about sell-offs…

Meta’s decision to eventually become a neocloud (a company that rents compute to others, like the other three Hyperscalers or CoreWeave) sent the stock up but crashed the entire market in return, despite Meta not being able to do so today.

The reported plan is for Meta to develop “Meta Compute,” a cloud-like business that would sell access to its data centers and AI models. Reuters says this could include access to hosted models, similar to AWS Bedrock, and possibly raw AI compute rental, similar to what neoclouds like CoreWeave sell. Meta declined to comment, and Reuters says it could not independently verify Bloomberg’s report.

The reason Vector’s headline says “excess compute it doesn’t have” is that “excess compute” normally means you built more capacity than you need internally. But Meta is simultaneously raising/maintaining enormous AI infrastructure spending, reportedly up to about $145 billion this year, which suggests the opposite: it needs far more GPUs/data centers/power for its own AI ambitions.

It’s interesting that a company with annual AI spending greater than Germany's defense spending defines the trigger for “Meta Compute” as having “excess compute.”

Meta’s investors loved the idea, but markets hated it because it once again unearthed the ongoing fear that the industry might be overbuilding. However, to be fair to Meta, weakness has already been present, and markets have been in a purgatory (not up, not down) for several weeks now.

Nonetheless, the main selling pressure came from chipmakers. Reuters reported that the Philadelphia semiconductor index fell 6.3% on July 1, dragging the Nasdaq down 0.66% and the S&P 500 down 0.22%. MarketWatch said the SOX index had already dropped 3.4% earlier in the session, after rising 87.8% in Q2, its strongest quarter on record.

The selloff also spread outside the US. The Economic Times reported that Samsung Electronics and SK Hynix fell as much as 14.5% on Thursday, while South Korea’s Kospi dropped 8.2%, with investors reacting to the same AI-capacity concerns.

In case you’re curious, knowing I’m an investor in both Hynix and Samsung and have heavy exposure to many other AI players, I haven’t sold.

But I do understand people’s fears and don’t blame them for doing so, as I myself explained to you recently how AI financing is concerning at best, really, really worrying at worst, especially when we factor in the growing presence of debt.

One thing is for a fully cash-and-equity-driven investment bubble to bust. When that happens, damage is contained. But once debt enters the picture, not only does risk spread across many more participants, but it also makes it almost impossible to know the extent of the exposure (e.g., shadow borrowing).

It’s fascinating, but as we saw, a data center project that fails to pay back a loan could prevent you from getting the money from your annuity your life insurer owes you because that money is now stuck in that failed project.

Those types of relationships are not remotely discussed in this industry, yet they are very real.

The Financial Times has published an article explaining that OpenAI has discussed giving the US government a 5% equity stake, worth roughly $42.6 billion, as part of talks to address political concerns around the AI industry and to share AI-related financial upside with the public.

The idea is described as early-stage and “conceptual.” Reports say it could require Congressional approval, and it may be part of a broader model in which the government would hold stakes in major US AI developers, potentially through a public wealth fund similar in concept to Alaska’s oil-funded dividend model.

It isn’t clear whether OpenAI expects the Administration to pour $42 billion in or if they are giving that stake for free. Either way, the intention is strikingly obvious to me: make OpenAI’s survival a state matter.

From the creators of “Too Big to Fall” comes “I’m an owner now; it can’t fail.” And while I understand why people like Bernie Sanders or Trump, among others, support the idea of the USG having a stake in these companies, it might end up being more of a bailout.

If OpenAI is worth 5 trillion one day, this will be seen as a massive success for the American people. But if things go south and OpenAI’s liquidity dries out, it will look more like a bailout than a national security investment.

Every day, a new example of how RLaaS (Reinforcement Learning as a Service, the idea of offering companies an easy way to train models) is going to take the world of AI by storm, emerges.

This time, it’s a partnership between Thinking Machines Labs (TML), the flash US AI Lab packed with ex-OpenAI and other top lab researchers that has bet its entire existence on this idea, and Bridgewater Associates, one of the world's largest hedge funds.

Using TML’s Tinker API, a product that lets you train open models with your own data without having to deal with the complexities of training and infrastructure, Bridgewater turns an open model (Chinese Qwen 3 235B, which is nowhere near the frontier) trained for filtering and processing financial documents into one that surfaces information relevant to investment decisions, into a state-of-the-art model, ahead of any commercially available model on the planet, as seen in the thumbnail.

Importantly, they also confirm that this model also beats heavily prompt-engineered frontier models, meaning they also tried to squeeze as much performance from AIs like Opus 4.8 or GPT-5-5 in an effort to see how far they can get with frontier models (which you can’t fine-tune), managing to reach high seventies results but still failing to beat the fine-tuned model, which beats them handsomely.

Also, as you can see in the thumbnail, it is not only a raw-score victory; the fine-tuned open model is between 12 and almost 20 times cheaper than the top OpenAI/Anthropic models.

Better performance at multiple times lower cost. Feels too good to be true, but that is why I’m so confident this is the only way to true enterprise adoption.

To make things even worse, they show how frontier models are getting more expensive than better, with GPT 5.4 costing 43% more than 5.2 but being only marginally more accurate.

And why is fine-tuning so superior to prompting? Well, as they explain: “Rather than contorting the expert’s intuition into a static prompt, the training process lets the model develop its own judgment.”

Following a pretty standard training pipeline, which of course includes on-policy distillation, which I talked about last week as key to AI progress these days, they convert a bad, open model into a frontier model (for that task).

You may see these results as unimpressive (“well, it’s only one task”), but here’s the thing: as I said above, enterprises don’t care about generalization to many tasks; they need the model to be great at that one task and will simply use another model for another task if required.

Now, while still surfing the foundational model wave, we are taking those models and making them great at single tasks again. This seems unavoidable to me, and the primary reason why Anthropic freaks out so much about open models.

As Weave Robotics posted, the company is launching Isaac 1, its home robot, with deliveries scheduled to begin this fall.

This appears to mark a shift from Isaac 0, a stationary laundry-folding robot, to Isaac 1, described as a more mobile home robot for a broader range of household tasks. A3’s earlier coverage said Isaac 0 was not the full mobile platform shown in Weave’s broader product vision, but a pared-down first deployment focused on laundry folding.

A related LinkedIn post from Weave co-founder Evan Wineland said the company had unveiled Isaac 1 at an event about two weeks earlier, and that Isaac 1’s design was recognized with an Industrial Design award by San Francisco Design Week.

The key thing here is that robotics is finally leaving the demo phase and actively looking to release products as soon as this year; it says orders are open and deliveries begin in fall 2026.

As for my personal opinion, I still struggle with the vision of having a humanoid in my home. Also, the helpful things it can really do for me aren't many; I don’t need a robot to make my bed.

People (pure coincidence, mostly investors) think this is going to change the world. But will it? I’m not sure.