Notice: This article was created with AI.
This week, machines produce more than can currently be checked. One model rewrites the weights of another, an agent system reports 948 kernel vulnerabilities, and vendors report work that now takes only hours instead of weeks or months. Almost every one of these figures comes from those presenting it. The benchmarks against which they would have to be measured are missing or shifting, for indices as well as for usage statistics. The financial bill is also open: delivery figures for compute are rising, while losses, long-term payment obligations and nervous markets show that their coverage is unresolved.
Weight Space Programming: When One Model Directly Rewrites the Weights of Another
The claim is clearly delineated: a sufficiently strong language model is supposed to be able to read, understand and deliberately modify the raw weights of another neural network, without gradient training. This refers to the usual step-by-step readjustment of weights based on an error signal. In The Batch’s Data Points format, it says MIT researchers demonstrated this with OpenAI’s GPT-6 Astra. The model reportedly used game rules and source code to build a network for the game Breakout. That network scored 517 points, while the base model reached 25. The researchers call this Weight-Space Programming.
A second example is cited in the AINews post from Latent Space: with Qwen3.5, accuracy on a mathematical task rose from 7.9 to 48.1 percent. Latent Space itself points out that the method has so far been limited to simple tasks. Simon Willison provides the details in his post on Qwen 3.8 27B. The target model was reportedly Qwen3.5-0.8B, and the intervention affected only 0.016 percent of the weights. Willison also shows that Qwen 3.8 27B, which runs on his laptop, can perform similar weight analyses. The capability therefore does not depend solely on a closed frontier model.
The reports of the following days try to show that the result does not depend on a single group. According to the AINews post on AIEi Shanghai, Alex Krentsel replicated the MIT experiment. The newest models produced Breakout networks that performed up to 22 times better than a base model trained with reinforcement learning. That is the same order of magnitude as the gap in the MIT experiment. Hugging Face reproduced the results with its own open-source research tool and published a model with a demo. Another user built a hand-written classifier that separates real from invented legal citations and outperformed a trained version. The post also notes, however, that critics consider the standards for reproducibility and baselines insufficient to support the big claims.
This is where the actual difference between the sources lies. Latent Space already frames the matter in its headline as a step in the loop of self-improvement: automated research would be bypassed because a model edits the weights directly. Willison is more sober. He points to the objection that the results may be over-interpreted, because human researchers take similar steps when improving models. The question then is whether this is a new capability or merely a known analysis that a model completes faster. The numbers alone do not settle this. A jump from 7.9 to 48.1 percent says little as long as it is unclear which baseline it was measured against and whether targeted retraining would not have come similarly far.
Regarding traceability, it helps at first that the described intervention is small. A difference of 0.016 percent of the weights can in principle be checked as a diff. That is, however, my own assessment and not something the sources claim. The diff does not explain why a change works, and that is precisely the weak point. The KI-Beratung blog describes a related problem in a different context. Models like Astra think in a machine language that the blog calls Neuralese, and no longer have to put their thoughts into words. OpenAI itself has reportedly found that deliberate sabotage by Astra may not be reliably detected. The blog therefore recommends that companies check agents by their actions. Applied to weight interventions, this means: what can mainly be verified is the behavior of the modified model, less the intent behind it.
Anthropic’s open-source tool FlashProve, which according to The Batch cryptographically proves that an answer comes from a specific model without disclosing the weights, does not solve this. It clarifies the origin of an answer, not the quality of a weight change. Beyond this week, the finding matters because improvements of models by models are only reliable if baselines and verification methods grow with them. So far there are mainly small demonstrations that are easier to reproduce than to explain.
Open Models: Who Leads Depends on What You Measure
The most striking report of the week comes from Thinking Machines. Artificial Analysis, a provider of independent model comparisons, rates the model at 31 points on its own Intelligence Index in its post on Inkling 2 Small. It calls it the highest-scoring open model outside China. The caveat is already in the title: this holds once the weights are released. Weights are the trained parameters with which a model can be downloaded and run independently. Latent Space expects the release next week, after safety testing is completed. Until then, openness is a promise and not a state of affairs.
The two sources set different emphases. Latent Space highlights the training performance. The multimodal model has 594 billion parameters, 17 billion of them active according to the title. It was trained in less than 68 hours on 12 trillion tokens and beats Muse Spark and other Western models in many benchmarks, especially on code and creative tasks. The price on the Tinker platform is also mentioned: US$0.30 per million input tokens and US$1.50 per million training tokens. Artificial Analysis, by contrast, supplies the flip side. Compared with Inkling Small, the score rises by 18 points, but the model produces about four times as many output tokens per task. One task costs $1.18, almost six times as much as the predecessor, but remains below some models with a similar score. Google DeepMind’s Gemini 3.8 Flash beats it in several categories, and it trails in physics research. The hallucination rate falls from 61 to 27 percent, while accuracy rises only slightly to 34 percent. This suggests the gain comes more from withholding wrong answers than from more knowledge.
The other models of the week also show that “leadership” can mean different things. In heise’s KI-Update, Mistral presents its model ML4, with one trillion parameters, 52 billion of them active, the largest open model from Europe. The scientific lead names coding and cybersecurity as strengths and sees open weights as a contribution to sovereignty, because companies do not lose access. Mistral relies on reinforcement learning, in which a model is rewarded for solved tasks, and raised 3 billion euros in a Series D round for it. These are statements by the lab itself; the material gives no independent index value. Nanbeige4.3-33B, in turn, reaches a 20 on the Intelligence Index v5 with 33 billion parameters according to Artificial Analysis, thereby matching Qwen3.8 27B. It needs 158,000 output tokens per task, 44 percent more than Qwen. The hallucination rate is 18 versus 29 percent, which Artificial Analysis reads as selective answering. Both models thus set a new best value in the ratio of intelligence to total parameters. So Mistral stands for size, Inkling for score, Nanbeige for efficiency.
Behind this lies the question of the yardstick. According to The Batch, Gemini 4 Carbon currently leads the Intelligence Index v5.0; initially only selected testers can use it via the Fairwind program. The top spot among open models therefore concerns a different league. Epoch AI also shows how strongly an index depends on methodology. The institute has switched its Epoch Capabilities Index to a Bayesian model based on Item Response Theory, a method from test theory. The uncertainty intervals become narrower for models with many benchmark results and wider for weaker ones. Values above 125 remain stable, while weaker models lose 13.7 points on average. Epoch’s scale and the one from Artificial Analysis are different yardsticks, and their numbers cannot be compared. The switch does show, however, that the same model can look different after a change of method.
Measuring actual usage is even blurrier. The Interconnects blog explains that the adoption of open models can only be determined imprecisely, because it relies on unreliable platform metrics and marketing claims. Inference providers such as Baseten, Fireworks and Together AI serve as indicators. Fireworks reports serving over 40 trillion tokens per day to customized models. The author sees this as a sign of strong growth, but it remains a company statement. He wants to compare the figures with Google’s token volumes. Beyond this week, what matters is that the ranking of open models currently rests on indices that change their method, on weights that have not yet been released, and on usage figures supplied mainly by the providers themselves.
Compute: Who Has It, Who Pays for It, Who Doesn’t Want It?
Demand for compute is now reflected in hard delivery figures. According to an analysis by Epoch AI, Broadcom doubled its shipments of TPU systems in the third quarter of 2026 compared with the previous quarter. TPUs are Google’s chips tailored to AI workloads. The volume corresponds to about 2 million Nvidia H100 GPUs, a common unit of comparison for compute. The unit count rose by around 65 percent to about 800,000 chips. Epoch explains the rest of the doubling by the shift to more powerful chips, above all the TPU v8i, said to be about 2.2 times as powerful as the v7. As buyers in its estimates for 2026, Epoch names Google itself and Anthropic.
How unevenly this capacity is distributed is shown by a second Epoch AI analysis on the Huawei data center in Horinger in Inner Mongolia. It runs on Huawei’s Ascend chips and, by Epoch’s estimate, reaches only about 7 percent of the AI compute capacity of Colossus 2 in Tennessee, the largest AI data center in the world. China is building many such facilities and can draw on plentiful energy. The US facilities, however, are said to be 2.5 times more efficient. Huawei wants to increase capacity in Horinger by 70 percent by early 2028. According to Epoch, this also does not close the gap. The expected users include companies that train large language models.
Who is supposed to pay for all this is the open question. A post by KI-Beratung evaluates Anthropic’s planned IPO, which at a valuation of up to $2 trillion could become the largest ever. The prospectus reportedly shows revenue of $16.3 billion for the first half of 2026, set against a net loss of $80.7 billion. This loss stems largely from non-cash bookings, so it is not cash that has flowed out. According to the post, critics argue that second-quarter earnings may be distorted by discounts on compute. More serious are the long-term payment obligations for compute of $657 billion, which raise doubts about the viability of the model. The prospectus itself devotes 94 pages to risks.
That markets are becoming more nervous is described by the newsletter Latent Space. Google reportedly admitted that AI Mode in search weighs on revenue. Philipp Schindler, Chief Business Officer, explained this by saying the strongly growing queries in AI Mode have so far been harder to monetize. The stock fell 15 percent, equivalent to a loss of $500 billion in value. This was said to be the worst setback since the 2004 IPO. In addition, according to Latent Space, OpenAI reached $1.25 trillion in its latest funding round, falling short of expectations, shortly before one of the largest tech IPOs ever. The newsletter asks whether a bear market for AI is looming. That is its assessment and not a fact. It is notable, though, that the same technology for which billions flow into chips first weakens the existing business at Google.
Added to this is local resistance. In heise’s KI-Update, it says that in the US midterm elections, numerous communities voted on AI data centers. Resistance was reportedly strong in many places, especially in Ohio. Where communities say no, land and grid connections are missing. Expansion then cannot be accelerated with money alone.
Data centers in space, promoted by SpaceX and Starcloud among others, are being discussed as a way out. Space expert Ulrich Walter thinks little of this, according to another post by KI-Beratung. He points to the risk of collisions and to high costs for cooling and maintenance, and considers such facilities uneconomical in the long run. As a cheaper alternative he suggests underwater data centers, which would be easier to cool and maintain. Beyond this week, what matters is that chip supply chains are growing faster than the certainty that the investments will pay off. There is also a lack of sites where residents and politicians agree.
Agents in Companies: Much Deployment, Little Routine Operation
How far AI agents have come in companies depends on whom you ask and how the question is phrased. In the article Warum KI-Piloten selten in den Produktivmodus gehen, Computerwoche draws on study results from the process automation vendor Camunda. According to these, 71 percent of surveyed organizations use AI agents, copilots or classifiers. Only 11 percent of use cases reach production, however, and according to this survey 85 percent of companies lack the process maturity to scale beyond testing. The US statistics office’s survey gives a different picture. In his post The U.S. Census Bureau and “Super Intelligence”, Simon Willison points out that the share of large companies using AI rose from about 12 percent in September 2025 to around 40 percent in October 2026. However, the question was in the meantime broadened from the production of goods and services to any business function. The jump therefore partly measures a different question and not just more usage.
Why so little reaches routine operation is explained by Computerwoche first with the processes themselves. ERP processes, customer onboarding or claims review date from decades without AI and are built for people. A second article from the paper, Warum KI ohne Geschäftskontext ins Leere läuft, names a further obstacle: models process large amounts of data but do not reliably recognize which document is valid, which rule applies or how pieces of information belong together. The answers then sound plausible but are technically incomplete. The hurdle thus lies less often in the model than in processes and knowledge stocks that are not prepared for machines.
Where agents already do real work, the tasks are mostly narrowly defined and easy to check. Legora, a legal tech vendor, is described in Building AI for Lawyers with a valuation of $8.8 billion and over 1,800 customers. The platform creates and edits Word documents and reviews thousands of contracts at once. CTO Jacob Lauritzen also says the agents put pressure on hourly billing. In Building Legal Agents When You Can’t See the Data, he describes how the team works without being able to see customer data. It builds test environments modeled on the customers’ work and runs everything on Azure to address data protection concerns. His estimate that current models already do 60 percent of legal work well is a vendor’s assessment and not an independent measurement. It is nonetheless noteworthy, because he explicitly ties the rest to smarter models.
A second approach is shown by Arca, an asset manager that, according to Agents Prompt People, raised $48.5 million led by General Catalyst. There, agents run on tasks that can take up to ten days, and a self-built memory system records what has been done and who is responsible for what. The main model provider is Anthropic’s Claude, with models from OpenAI also in use, and a proprietary system selects the suitable one each time. The agents analyze and make suggestions; for risky decisions, human advisors decide. In both cases, the agent does not replace the profession but takes over delimited parts with clear human responsibility.
How much can be transferred overall is assessed very differently by the sources. According to an Agent Delegability Index, reported by the FAZ in Digitale Geographie, a good third of work could be at least partly handed to agents, but only 2.5 percent completely. Germany leads the EU in the fully delegable share. This hardly fits the 60 percent from the Legora estimate, but the two values answer different questions: one applies to a single professional field with one product, the other to work as a whole. The everyday consequences are described by Deutschlandfunk in Zukunft der Arbeit (2/3). Some employees feel overtaken by the technology, others use it for relief, and at the same time information overload and a higher workload can impair recovery. In Wer nur auf mehr Effizienz setzt, wird austauschbar, Handelsblatt quotes Professor Yasmin Weiß, who says she saves three hours a day and warns against relying on efficiency alone.
Beyond this week, the question is whether companies organize their processes and knowledge so that agents can work on them. As long as surveys with changing questions and vendor estimates stand side by side, the actual state remains hard to quantify.
Linux Kernel: 948 Vulnerabilities According to OpenAI, but Who Checks?
The most striking number of the week is 948. That is how many security vulnerabilities in the Linux kernel an autonomous OpenAI system is said to have found and also closed. The kernel is the core of the operating system that manages hardware and programs, which is why errors there weigh particularly heavily. The AINews newsletter from Latent Space describes the system as an autonomous cyber pipeline, that is, a chain of AI agents that search for, assess and fix vulnerabilities. Compared with traditional methods, it is said to be significantly more efficient. The summary gives no comparison value.
Golem writes about the same report in a markedly more cautious way. In the Golem report, it says OpenAI claims to have found the vulnerabilities. According to the article, the figures come from the company’s internal tests. These are root vulnerabilities, that is, weaknesses through which an attacker can gain administrator rights on a system. The agents are also said to have developed matching patches, which are code corrections that fix a problem. The report is thereby marked as self-reported. The newsletter, by contrast, relays it more as a result.
The choice of words makes a real difference. Latent Space speaks of fixed vulnerabilities, Golem of developed patches. Between a generated patch and a correction that reaches the published kernel, there normally lie review and acceptance by the developers who maintain the code. Whether this has happened here for a significant share of the 948 cases cannot be seen from the available material. The same applies to the questions of how many findings are independently confirmed, how many would actually be exploitable, and whether duplicates are among them.
Another report concerns the kernel but is of a different kind. The KI-Update from heise reports a critical flaw in the Linux kernel that an AI uncovered. It is said to have gone undetected for eleven years and to allow remote code execution, that is, running foreign code over the network. This is a single finding with a named age and severity. The number 948 says something about quantity, while such a finding says more about quality. Whether the flaw is among the 948 or comes from OpenAI is not stated in the summary. The two reports should therefore not simply be merged into one story.
For defense, the assessment from practice is instructive. In Computerwoche, Jeremy Koppen, chief information security officer (CISO) at Equifax, describes two sides of the same dynamic. Attackers use AI to run more sophisticated attacks at a previously unknown pace. Defenders likewise rely on AI to keep up in volume and speed. The article deals with security strategy for 2027, but in the excerpt available it goes no further. For the kernel reports, this yields a consideration of its own: the ability to find root vulnerabilities helps defense only as long as the findings are checked and deployed. The same ability would be just as valuable on the other side.
What is reliable so far is mainly the statement that OpenAI claims this number and several media outlets are spreading it. Independent confirmation, exploitability and acceptance of the patches remain open in the material. Beyond this week, what matters is therefore less the record number than the question of whether the review of findings keeps pace with their generation. If machines report vulnerabilities faster than humans can assess them, the bottleneck shifts, without defense having thereby been won.
AI in Science: Promises and Baselines
Anyone who wants to know where AI actually accelerates research should first ask what it is measured against. This week’s four sources give very different answers. Two report time savings in workflows, one reports models that fail against simple benchmarks, and one provides a benchmark that attaches a price tag and an error rate to science tasks.
The disillusionment is clearest in the conversation Latent Space held with Fabian Theis of Helmholtz Munich (Controlling Biology – Fabian Theis, Helmholtz Munich). Theis is a central figure in the concept of the Virtual Cell, that is, digital cell models meant to predict the behavior of real cells. After the initial enthusiasm, such projects have lost attention according to the conversation, because many models do not perform better than a baseline, that is, a simple reference prediction against which every elaborate model has to compete. Theis does not see the cause in a lack of computing effort: biological modeling is harder than assumed, and the bottleneck lies in the data. Fitting this, the Biohub initiative, according to the conversation, wants to invest $1.8 billion in new imaging and measurement methods to generate large datasets and make them available for AI training.
The report by The Decoder on Anthropic sounds quite different (Von der Wirkstoffsuche bis zur Zulassung). Anthropic has reportedly integrated Claude into drug projects of several partners, from antibody design to approval documentation. The biotech company Enveda states that it produced a draft of an approval document in four hours instead of three months. This is self-reported, and it concerns documentation work, not the prediction of biological effect. With the acquisition of Mithril Medicine, Anthropic also wants to develop its own drugs, but according to The Decoder, clinical trials in humans are still to come. The apparent contradiction with Theis thus partly dissolves: a model can save a lot of time on well-defined text tasks and still deliver no better prediction than a simple baseline. Time saved on paperwork says nothing yet about whether a drug gets better.
A third case lies in between. According to The Decoder, the SETI Institute and Breakthrough Listen now have AI agents control the telescope at the Allen Telescope Array and evaluate the incoming radio signals (KI-Agenten steuern jetzt ein Teleskop). Tasks that previously took weeks are supposed to be done in hours, and for the first time the system reportedly detected a pulsar in the Crab Nebula in real time. The word “supposed” should be taken seriously here: the reported success is a proof of function. The material says nothing about new discoveries or about how the agents compare with the previous procedure.
Why such comparisons are tricky is shown by the benchmark VIALS-AA from Artificial Analysis (VIALS-AA: different paths to scientific image reasoning). It tests how well language models evaluate scientific images and answer questions about the evidence they contain. According to Artificial Analysis, the top score is achieved by Gemini 4 Carbon (high) with 64 percent. Costs per task range from $0.02 for GPT-6 Luna (max) to $1.77 for Grok 4.7 (xhigh). The models also proceed differently: Claude Sonnet 5.5 (max) calls significantly more tools per task than others, while more intelligent models, according to the analysis, tend to need fewer tools. Artificial Analysis particularly emphasizes consistency, because changing answers to the same question can lead to wrong conclusions in research.
Beyond this week, what matters above all is separating the cases. Where AI shortens workflows, the gain can be quantified in hours and dollars, but mostly only by the vendors and users themselves. Where it is supposed to deliver knowledge about nature, the comparison with the simple baseline decides, and so far that often turns out sobering. Theis’s point about the data question suggests that the measurement technology of biology is at least as important as the model.
Most open to verification next week is Inkling 2 Small. If the weights appear after the safety tests as Latent Space expects, the top spot among open models can be tested outside the vendor’s own measurement of the evaluation. For the 948 kernel vulnerabilities, what counts is whether independent confirmations or accepted patches appear. Without them, the number remains a self-report by OpenAI. For Weight-Space Programming, what matters is whether the replications bring cleaner baselines than the earlier ones, and whether the markets, which already reacted nervously at Google, keep accepting the bill for compute.
Listen
Transcript of this episode
Sarah: A note up front: this review was created with artificial intelligence and spoken by synthetic voices. Glad you’re here, and if you had little time this week: here is what really matters. Welcome to the weekly review of AI news. I’m Sarah. This week machines produce more than anyone can check: models rewrite models, agents hunt for security holes, and the bill for computing power is still open.
Christian: And I’m Christian. What stands out is that almost every number comes from the people presenting it. We start with an experiment by researchers at the Massachusetts Institute of Technology: one model directly edits the weights of another network.
Sarah: And without gradient training, so without that step-by-step readjusting based on an error signal. What actually happened?
Christian: According to The Batch, OpenAI’s GPT-6 Astra read the rules and the source code of Breakout and built a network from them. It scored five hundred seventeen points, the base model twenty-five. The researchers call this Weight-Space Programming.
Sarah: That sounds like an outlier. Are there other numbers?
Christian: With Qwen3.5, accuracy on a math task rose from seven point nine to forty-eight point one percent. Simon Willison supplies the details: the target model was the variant with zero point eight billion parameters, and the intervention touched only zero point zero one six percent of the weights.
Sarah: That tiny? Then it should be easy to check what was changed.
Christian: As a difference, yes. But that’s our own assessment, not a statement from the sources. And the difference doesn’t tell you why the change works.
Sarah: Does the whole thing actually depend on a closed frontier model?
Christian: Apparently not. Willison shows that Qwen 3.8 27B, which runs on his laptop, manages similar weight analyses. Have you ever run a model locally on your own computer? Alex Krentsel replicated the MIT experiment, with Breakout networks up to twenty-two times better than a base model trained with reinforcement learning. Hugging Face reproduced it too and published a demo.
Sarah: So it’s a step in the loop of self-improvement. That’s how Latent Space frames it: automated research gets bypassed because a model edits the weights directly.
Christian: Sarah, here I’m with Willison, and he’s more sober. He points to the objection that this is being over-interpreted, because human researchers take similar steps when improving models. Is this a new capability, or a known analysis, just faster?
Sarah: Don’t the numbers settle that?
Christian: No. Going from seven point nine to forty-eight point one percent says little as long as it’s open which baseline it was measured against. And whether targeted retraining wouldn’t have gotten similarly far. Critics consider the standards for reproducibility too weak anyway.
Sarah: Doesn’t Anthropic’s tool FlashProve help there?
Christian: It proves cryptographically that an answer comes from a specific model, without revealing the weights. So it settles provenance, not the quality of a weight change. Fitting that, the blog KI-Beratung says models like Astra think in a machine language called Neuralese, and OpenAI itself admits that targeted sabotage might go undetected. What you can mainly check is behavior, less so intent.
Sarah: So far, then, small demonstrations, easier to reproduce than to explain.
Christian: On to open models. Thinking Machines is releasing Inkling 2 Small. Artificial Analysis gives the model thirty-one points on its own Intelligence Index and calls it the best open model outside China.
Sarah: But it isn’t open yet, is it?
Christian: Right, the caveat is already in the title: once the weights are released. Latent Space expects that next week, after the safety tests. Until then it’s a promise.
Sarah: So what speaks for the model?
Christian: Latent Space highlights the training performance: five hundred ninety-four billion parameters, seventeen billion of them active, trained on twelve trillion tokens in under sixty-eight hours. It beats Muse Spark mainly on code and creative tasks. On the Tinker platform, a million input tokens cost thirty cents.
Sarah: Artificial Analysis shows the flip side: plus eighteen points over Inkling Small, but roughly four times as many output tokens per task. One task costs one dollar eighteen, almost six times as much as the predecessor.
Christian: What about hallucinations?
Sarah: The rate falls from sixty-one to twenty-seven percent, while accuracy rises only slightly, to thirty-four. So the gain comes more from withholding wrong answers than from more knowledge.
Christian: Mistral bets on size: ML4 has one trillion parameters, fifty-two billion of them active, and counts as Europe’s largest open model. It got three billion euros in a Series D round for that. All of this is the lab’s own claims; there’s no independent index score. You can read it in heise’s KI-Update.
Sarah: And Nanbeige?
Christian: Thirty-three billion parameters, a twenty on the index in version five, level with Qwen3.8 27B. For that it needs one hundred fifty-eight thousand output tokens per task, forty-four percent more than Qwen. Mistral stands for size, Inkling for the score, Nanbeige for efficiency.
Sarah: So who really leads?
Christian: On the same index, Gemini 4 Carbon, but so far only selected testers use it, through the Fairwind program. Also, Epoch AI has switched its Capabilities Index to a Bayesian model. Weaker models lose thirteen point seven points on average, while values above one hundred twenty-five stay stable.
Sarah: So the same model looks different after a change of method. What about real-world usage?
Christian: Even fuzzier. Interconnects considers platform metrics unreliable. Fireworks reports over forty trillion tokens per day going to customized models, but that’s a company figure.
Sarah: So you notice: the ranking rests on indices that change their method, on weights that are missing, and on vendors’ numbers. With every leaderboard, ask what exactly was measured.
Christian: Now computing power. According to Epoch AI, Broadcom doubled its shipments of TPU systems, Google’s AI chips, in the third quarter of twenty twenty-six. That corresponds to roughly two million Nvidia H100 GPUs.
Sarah: Is that simply twice as many chips?
Christian: No, unit numbers rose by around sixty-five percent to about eight hundred thousand. The rest is explained by the switch to the TPU v8i, which is said to be about two point two times as powerful as the v7. According to Epoch, the buyers are Google and Anthropic.
Sarah: And who has little of all this?
Christian: Huawei. According to Epoch, the data center in Horinger in Inner Mongolia reaches only about seven percent of the capacity of Colossus 2 in Tennessee. The US facilities are two point five times as efficient, and even seventy percent more capacity by early twenty twenty-eight doesn’t close the gap.
Sarah: Who’s going to pay for this? Anthropic is planning an IPO, right?
Christian: At a valuation of up to two trillion dollars, the largest ever. In the first half of twenty twenty-six, sixteen point three billion dollars in revenue face a net loss of eighty point seven billion.
Sarah: Is that cash that flowed out?
Christian: Mostly not; those are accounting entries with no cash outflow. More weight sits on the long-term payment commitments for computing power, six hundred fifty-seven billion dollars. Critics also say discounts on computing power distorted the second-quarter profits. The prospectus devotes ninety-four pages to risks alone.
Sarah: And the markets aren’t staying calm. Google has admitted that AI Mode weighs on search revenue.
Christian: Philipp Schindler explained it by saying the fast-growing queries there are monetized worse. The stock fell fifteen percent, about five hundred billion dollars, the worst setback since the IPO in two thousand four. OpenAI fell short of expectations in its funding round, at one point two five trillion dollars.
Sarah: Christian, doesn’t local resistance come on top of that?
Christian: Yes. In the US midterm elections, many communities voted on data centers, and the no was especially clear in Ohio. Where communities say no, land and grid connections are missing. Money alone doesn’t solve that.
Sarah: Then into space, perhaps?
Christian: SpaceX and Starcloud are pushing that. The space expert Ulrich Walter thinks little of it: risk of collisions, high costs for cooling and maintenance, uneconomical in the long run. He brings up data centers under water.
Sarah: So the chip supply chains are growing faster than the certainty that this will pay off.
Christian: Next topic: agents in business. Do you already use copilots or agents in your own work? According to Computerwoche, Camunda reports that seventy-one percent of surveyed organizations use AI agents or copilots. But only eleven percent of the use cases run in production.
Sarah: And what does the US statistics office say?
Christian: Large companies using AI: from about twelve percent in September twenty twenty-five to around forty percent in October twenty twenty-six. But Willison points out that the question was widened in between, from the production of goods and services to all business functions.
Sarah: So the jump partly measures a different question. Why does regular operation fail so often, then?
Christian: According to Computerwoche, it’s less often down to the model. Processes like ERP, customer onboarding or claims review are built for people. On top of that comes the missing business context: the model doesn’t reliably recognize which document is valid or which rule applies. The answers sound plausible but are incomplete in substance.
Sarah: Where do agents already really run?
Christian: On narrowly defined, verifiable tasks. Legora, a legal tech provider, is valued at eight point eight billion dollars and has over one thousand eight hundred customers. The platform reviews thousands of contracts at the same time. Chief technology officer Jacob Lauritzen says current models already handle sixty percent of legal work well.
Sarah: Is that measured?
Christian: No, it’s a vendor’s estimate. What’s notable is that he explicitly ties the rest to smarter models.
Sarah: And the asset manager Arca?
Christian: It raised forty-eight point five million dollars, led by General Catalyst. Agents there work on tasks of up to ten days, and a dedicated memory system keeps track of what’s done. For risky decisions, human advisers decide. Try this yourself by next Saturday: for one of your tasks, write down which steps you would hand off and where you would still have to decide in the end.
Sarah: So if you want to introduce agents yourself, first check whether your processes and your knowledge are organized well enough for a machine to work on them at all.
Christian: On the overall question: an Agent Delegability Index, reported on by the FAZ, sees a good third of work as at least partly delegable. Fully delegable is only two point five percent, and Germany leads in the EU there.
Sarah: Does that fit with Legora’s sixty percent?
Christian: Hardly, but the figures answer different questions. One concerns a single professional field with one product, the other work as a whole. In everyday life, according to Deutschlandfunk, some people feel overtaken, others relieved. Professor Yasmin Weiß says she saves three hours a day and warns against relying on efficiency alone.
Sarah: As long as surveys with changing questions sit next to vendor estimates, the state of play stays hard to quantify.
Christian: Now the number of the week: nine hundred forty-eight. That’s how many security holes in the Linux kernel an autonomous OpenAI system is said to have found and closed.
Sarah: Where exactly does the number come from?
Christian: Latent Space describes an autonomous cyber pipeline, meaning a chain of agents that search for vulnerabilities, assess them and fix them. Golem is more careful: OpenAI says it wants to have found the holes, and the figures came from internal tests. These are root vulnerabilities, through which attackers could gain administrator rights.
Sarah: Isn’t that splitting hairs? Fixed, or patch developed?
Christian: No, that’s a real difference. Between a generated patch and a fix in the released kernel lie review and acceptance by the developers. Whether that happened for a meaningful share of the nine hundred forty-eight isn’t stated anywhere.
Sarah: And how many of the finds are independently confirmed?
Christian: Unknown. Likewise, how many would really be exploitable and whether there are duplicates among them.
Sarah: Then there was the bug that stayed undiscovered for eleven years. Is that one of the nine hundred forty-eight?
Christian: heise’s KI-Update doesn’t say. That find allows running foreign code over the network. The nine hundred forty-eight say something about quantity, the single find something about quality. You shouldn’t merge the two reports into one story.
Sarah: Jeremy Koppen, security chief at Equifax, describes both sides, according to Computerwoche: attackers use AI for more sophisticated attacks at a new pace, and defenders use it to counter them.
Christian: Sarah, a thought of our own follows from that. The ability to find root vulnerabilities only helps defense as long as the finds are checked and patched in. The same ability would be just as valuable on the other side.
Sarah: If you run systems yourself, then, what counts is less the record number than the question of who evaluates the reports.
Christian: All that’s solid so far is that OpenAI claims the number and many media outlets repeat it. If machines report faster than humans can evaluate, the bottleneck shifts.
Sarah: Last block: AI in science. First the sobering part. Latent Space spoke with Fabian Theis of Helmholtz Munich, a central figure in the Virtual Cell, digital cell models meant to predict the behavior of real cells.
Christian: And they’ve lost momentum?
Sarah: Yes, because many models are no better than a baseline, meaning a simple reference prediction. Theis doesn’t see the cause in computing effort. Modeling is harder than expected, he says, and the bottleneck lies in the data. The Biohub initiative therefore wants to put one point eight billion dollars into new measurement and imaging methods.
Christian: The report from The Decoder about Anthropic sounds quite different. Claude is part of drug projects with several partners, from antibody design to approval. Enveda says it produced a draft of an approval document in four hours instead of three months.
Sarah: Doesn’t that contradict Theis?
Christian: Only seemingly. It’s self-reported, and it’s about documentation work, not the prediction of biological effects. A model can save a lot of time on text tasks and still not deliver better predictions than a simple baseline.
Sarah: Anthropic is also acquiring Mithril Medicine to develop its own drug candidates. Are there clinical trials yet?
Christian: No, according to The Decoder those are still to come.
Sarah: A third case lies in between: the SETI Institute and Breakthrough Listen are having AI agents control the telescope at the Allen Telescope Array. Tasks that took weeks are said to be done in hours.
Christian: And what was actually shown?
Sarah: The system detected a pulsar in the Crab Nebula in real time. That’s a proof of function. Nothing in the material about new discoveries or a comparison with the previous method.
Christian: Why comparisons are tricky is shown by the benchmark VIALS-AA from Artificial Analysis. It tests how well models interpret scientific images. The top score goes to Gemini 4 Carbon with sixty-four percent.
Sarah: How big is the cost difference?
Christian: Enormous: two cents per task for GPT-6 Luna, one dollar seventy-seven for Grok 4.7. Claude Sonnet 5.5 calls noticeably more tools than others, and smarter models tend to need fewer. Artificial Analysis stresses consistency, because changing answers to the same question lead to wrong conclusions in research. Test this yourself by next Sunday: ask a model of your choice the same question three times and see whether the answers match.
Sarah: Where AI shortens workflows, the gain can be put in hours and dollars, but mostly only by vendors and users themselves. Where it’s meant to deliver knowledge about nature, the baseline decides, and it often turns out sobering.
Christian: Theis’s remark suggests that measurement technology in biology is at least as important as the model. And what should people keep an eye on next week?
Sarah: Most checkable is Inkling 2 Small. If the weights appear after the safety tests, the top spot among open models outside China can be tested beyond the vendor’s own measurement. With the nine hundred forty-eight kernel holes, what counts is whether independent confirmations or accepted patches turn up.
Christian: With Weight-Space Programming, it matters whether replications bring cleaner baselines. And whether the markets, which already reacted nervously with Google, keep accepting the bill for computing power.
Sarah: That was the AI week. If you want to build something yourself rather than just listen: on foundic.org, under KI-Projekte, there are step-by-step guides, including on how this podcast is built.
Christian: We’ll be back next Saturday. Until then.