Anthropic’s latest research suggests that while AI is rapidly changing the way work gets done, it hasn’t meaningfully eliminated jobs. At least, not yet. But beneath what Anthropic’s head of economics, Peter McCrory, says is a “still healthy” labor market, early signs are pointing to uneven impacts, especially for younger workers just entering the workforce. 

In an interview on the sidelines of the Axios AI Summit in Washington, D.C., McCrory said the company’s newest economic impact report finds little evidence of widespread job displacement so far. 

“There’s no material difference in unemployment rates” between workers who use Claude for the “most central task of their job in automated ways” — like technical writers, data entry clerks, and software engineers — and workers in jobs less exposed to AI that require “physical interaction and dexterity with the real world.” 

But with AI adoption spreading across industries, that could shift — fast. If Anthropic CEO Dario Amodei is to be believed, AI could wipe out half of all entry-level white-collar jobs and push unemployment as high as 20% within the next five years.

“Displacement effects could materialize very quickly, so you want to establish a monitoring framework to understand that before it materializes so that we can catch it as it’s happening and ideally identify the appropriate policy response,” McCrory told TechCrunch.

Staying ahead of those trends is why tracking AI growth, adoption, and diffusion is so important, he said.

In theory, McCrory said, AI models like Claude can do almost anything a computer can do. In practice, most users are only scratching the surface of those capabilities.

He said Anthropic looked at which roles involve tasks that AI is particularly good at, that are already being automated, and that are tied to real workplace use cases — the areas most likely to signal where displacement could emerge. 

Anthropic’s fifth economic impact report, released Tuesday, also found that even where there hasn’t been much displacement yet, there’s a growing skills gap between earlier Claude adopters and newcomers.

Earlier adopters are more likely to get significantly more value from the model, using it for work-related tasks rather than casual or one-off purposes and in more sophisticated ways, like as a “thought partner” for iteration and feedback. 

McCrory said the findings suggest AI is becoming a technology that rewards those who already know how to use it — and that workers who can effectively incorporate it into their work will increasingly have an edge.

That advantage isn’t evenly distributed geographically, either. The report also found that “Claude is used more intensely in high-income countries, within the U.S. in places with more knowledge workers, and for a relatively small set of specialized tasks and occupations.”

In other words, despite promises of AI as an equalizer, adoption may already be tilting toward the wealthy and could amplify those advantages as power users pull further ahead.

Imagine the perfect online shop. It’d offer great deals on the biggest tech, gaming and entertainment brands. It’d give you same-day delivery without charging extra. It’d have real humans answering the phone and 24/7 customer service. And it would stock everything from AirPods and action cameras to air fryers and large appliances.

We’ve just described Joybuy, a fantastic new place to shop for almost anything – and to celebrate its UK launch it’s offering amazing launch deals including up to 50% off selected items from big brands, a “spend £99 and save £10” on selected products offer and a “spend £200 and save £100” deal on selected home appliances. And while we’re interested in the amazing tech deals you’ll also be able to get some deep discounts on appliances, beauty, groceries and more.

The names you know at prices you’ll love

As Large Language Models (LLMs) expand their context windows to process massive documents and intricate conversations, they encounter a brutal hardware reality known as the “Key-Value (KV) cache bottleneck.”

Every word a model processes must be stored as a high-dimensional vector in high-speed memory. For long-form tasks, this “digital cheat sheet” swells rapidly, devouring the graphics processing unit (GPU) video random access memory (VRAM) system used during inference, and slowing the model performance down rapidly over time.

But have no fear, Google Research is here: yesterday, the unit within the search giant released its TurboQuant algorithm suite — a software-only breakthrough that provides the mathematical blueprint for extreme KV cache compression, enabling a 6x reduction on average in the amount of KV memory a given model uses, and 8x performance increase in computing attention logits, which could reduce costs for enterprises that implement it on their models by more than 50%.

The theoretically grounded algorithms and associated research papers are available now publicly for free, including for enterprise usage, offering a training-free solution to reduce model size without sacrificing intelligence.

The arrival of TurboQuant is the culmination of a multi-year research arc that began in 2024. While the underlying mathematical frameworks—including PolarQuant and Quantized Johnson-Lindenstrauss (QJL)—were documented in early 2025, their formal unveiling today marks a transition from academic theory to large-scale production reality.

The timing is strategic, coinciding with the upcoming presentations of these findings at the upcoming conferences International Conference on Learning Representations (ICLR 2026) in Rio de Janeiro, Brazil, and Annual Conference on Artificial Intelligence and Statistics (AISTATS 2026) in Tangier, Morocco.

By releasing these methodologies under an open research framework, Google is providing the essential “plumbing” for the burgeoning “Agentic AI” era: the need for massive, efficient, and searchable vectorized memory that can finally run on the hardware users already own. Already, it is believed to have an effect on the stock market, lowering the price of memory providers as traders look to the release as a sign that less memory will be needed (perhaps incorrect, given Jevons’ Paradox).

The Architecture of Memory: Solving the Efficiency Tax

To understand why TurboQuant matters, one must first understand the “memory tax” of modern AI. Traditional vector quantization has historically been a “leaky” process.

When high-precision decimals are compressed into simple integers, the resulting “quantization error” accumulates, eventually causing models to hallucinate or lose semantic coherence.

Furthermore, most existing methods require “quantization constants”—meta-data stored alongside the compressed bits to tell the model how to decompress them. In many cases, these constants add so much overhead—sometimes 1 to 2 bits per number—that they negate the gains of compression entirely.

TurboQuant resolves this paradox through a two-stage mathematical shield. The first stage utilizes PolarQuant, which reimagines how we map high-dimensional space.

Rather than using standard Cartesian coordinates (X, Y, Z), PolarQuant converts vectors into polar coordinates consisting of a radius and a set of angles.

The breakthrough lies in the geometry: after a random rotation, the distribution of these angles becomes highly predictable and concentrated. Because the “shape” of the data is now known, the system no longer needs to store expensive normalization constants for every data block. It simply maps the data onto a fixed, circular grid, eliminating the overhead that traditional methods must carry.

The second stage acts as a mathematical error-checker. Even with the efficiency of PolarQuant, a residual amount of error remains. TurboQuant applies a 1-bit Quantized Johnson-Lindenstrauss (QJL) transform to this leftover data. By reducing each error number to a simple sign bit (+1 or -1), QJL serves as a zero-bias estimator. This ensures that when the model calculates an “attention score”—the vital process of deciding which words in a prompt are most relevant—the compressed version remains statistically identical to the high-precision original.

Performance benchmarks and real-world reliability

The true test of any compression algorithm is the “Needle-in-a-Haystack” benchmark, which evaluates whether an AI can find a single specific sentence hidden within 100,000 words.

In testing across open-source models like Llama-3.1-8B and Mistral-7B, TurboQuant achieved perfect recall scores, mirroring the performance of uncompressed models while reducing the KV cache memory footprint by a factor of at least 6x.

This “quality neutrality” is rare in the world of extreme quantization, where 3-bit systems usually suffer from significant logic degradation.

Beyond chatbots, TurboQuant is transformative for high-dimensional search. Modern search engines increasingly rely on “semantic search,” comparing the meanings of billions of vectors rather than just matching keywords. TurboQuant consistently achieves superior recall ratios compared to existing state-of-the-art methods like RabbiQ and Product Quantization (PQ), all while requiring virtually zero indexing time.

This makes it an ideal candidate for real-time applications where data is constantly being added to a database and must be searchable immediately. Furthermore, on hardware like NVIDIA H100 accelerators, TurboQuant’s 4-bit implementation achieved an 8x performance boost in computing attention logs, a critical speedup for real-world deployments.

Rapt community reaction

The reaction on X, obtained via a Grok search, included a mixture of technical awe and immediate practical experimentation.

The original announcement from @GoogleResearch generated massive engagement, with over 7.7 million views, signaling that the industry was hungry for a solution to the memory crisis.

Within 24 hours of the release, community members began porting the algorithm to popular local AI libraries like MLX for Apple Silicon and llama.cpp.

Technical analyst @Prince_Canuma shared one of the most compelling early benchmarks, implementing TurboQuant in MLX to test the Qwen3.5-35B model.

Across context lengths ranging from 8.5K to 64K tokens, he reported a 100% exact match at every quantization level, noting that 2.5-bit TurboQuant reduced the KV cache by nearly 5x with zero accuracy loss. This real-world validation echoed Google’s internal research, proving that the algorithm’s benefits translate seamlessly to third-party models.

Other users focused on the democratization of high-performance AI. @NoahEpstein_ provided a plain-English breakdown, arguing that TurboQuant significantly narrows the gap between free local AI and expensive cloud subscriptions.

He noted that models running locally on consumer hardware like a Mac Mini “just got dramatically better,” enabling 100,000-token conversations without the typical quality degradation.

Similarly, @PrajwalTomar_ highlighted the security and speed benefits of running “insane AI models locally for free,” expressing “huge respect” for Google’s decision to share the research rather than keeping it proprietary.

Market impact and the future of hardware

The release of TurboQuant has already begun to ripple through the broader tech economy. Following the announcement on Tuesday, analysts observed a downward trend in the stock prices of major memory suppliers, including Micron and Western Digital.

The market’s reaction reflects a realization that if AI giants can compress their memory requirements by a factor of six through software alone, the insatiable demand for High Bandwidth Memory (HBM) may be tempered by algorithmic efficiency.

As we move deeper into 2026, the arrival of TurboQuant suggests that the next era of AI progress will be defined as much by mathematical elegance as by brute force. By redefining efficiency through extreme compression, Google is enabling “smarter memory movement” for multi-step agents and dense retrieval pipelines. The industry is shifting from a focus on “bigger models” to “better memory,” a change that could lower AI serving costs globally.

Strategic considerations for enterprise decision-makers

For enterprises currently using or fine-tuning their own AI models, the release of TurboQuant offers a rare opportunity for immediate operational improvement.

Unlike many AI breakthroughs that require costly retraining or specialized datasets, TurboQuant is training-free and data-oblivious.

This means organizations can apply these quantization techniques to their existing fine-tuned models—whether they are based on Llama, Mistral, or Google’s own Gemma—to realize immediate memory savings and speedups without risking the specialized performance they have worked to build.

From a practical standpoint, enterprise IT and DevOps teams should consider the following steps to integrate this research into their operations:

Optimize Inference Pipelines: Integrating TurboQuant into production inference servers can reduce the number of GPUs required to serve long-context applications, potentially slashing cloud compute costs by 50% or more.

Expand Context Capabilities: Enterprises working with massive internal documentation can now offer much longer context windows for retrieval-augmented generation (RAG) tasks without the massive VRAM overhead that previously made such features cost-prohibitive.

Enhance Local Deployments: For organizations with strict data privacy requirements, TurboQuant makes it feasible to run highly capable, large-scale models on on-premise hardware or edge devices that were previously insufficient for 32-bit or even 8-bit model weights.

Re-evaluate Hardware Procurement: Before investing in massive HBM-heavy GPU clusters, operations leaders should assess how much of their bottleneck can be resolved through these software-driven efficiency gains.

Ultimately, TurboQuant proves that the limit of AI isn’t just how many transistors we can cram onto a chip, but how elegantly we can translate the infinite complexity of information into the finite space of a digital bit. For the enterprise, this is more than just a research paper; it is a tactical unlock that turns existing hardware into a significantly more powerful asset.

Attacks leveraging the ‘PolyShell’ vulnerability in version 2 of Magento Open Source and Adobe Commerce installations are underway, targeting more than half of all vulnerable stores.

According to eCommerce security company Sansec, hackers started exploiting the critical PolyShell issue en masse last week, just two days after public disclosure.

“Mass exploitation of PolyShell started on March 19th, and Sansec has now found PolyShell attacks on 56.7% of all vulnerable stores,” Sansec says.

The researchers previously reported that the problem lies in Magento’s REST API, which accepts file uploads as part of the custom options for the cart item, allowing polyglot files to achieve remote code execution or account takeover via stored cross-site scripting (XSS), if the web server configuration allows it.

Adobe released a fix in version 2.4.9-beta1 on March 10, 2026, but it has not yet reached the stable branch. BleepingComputer previously contacted Adobe to ask about when a security update addressing PolyShell will become available for production versions, but we have not received a response.

Meanwhile, Sansec has published a list of IP addresses that target scanning for web stores vulnerable to PolyShell.

WebRTC skimmer

Sansec reports that in some of the attacks suspected to exploit PolyShell, the threat actor delivers a novel payment card skimmer that uses Web Real-Time Communication (WebRTC) to exfiltrates data.

WebRTC uses DTLS-encrypted UDP rather than HTTP, so it is more likely to evade security controls even on sites with strict Content Security Policy (CSP) controls like “connect-src.”

The skimmer is a lightweight JavaScript loader that connects to a hardcoded command-and-control (C2) server via WebRTC, bypassing normal signaling by embedding a forged SDP exchange.

It receives a second-stage payload over the encrypted channel, then executes it while bypassing CSP, primarily by reusing an existing script nonce, or falling back to unsafe-eval or direct script injection. Execution is delayed using ‘requestIdleCallback’ to reduce detection.

Sansec noted that this skimmer was detected on the e-commerce website of a car maker valued at over $100 billion, which did not respond to their notifications.

The researchers provide a set of indicators of compromise that can help defenders protect against these attacks.

Malware is getting smarter. The Red Report 2026 reveals how new threats use math to detect sandboxes and hide in plain sight.

Download our analysis of 1.1 million malicious samples to uncover the top 10 techniques and see if your security stack is blinded.

Download The Report

from the melted-snowflakes dept

Last week we covered how the government successfully convinced Judge Colleen McMahon to order the plaintiffs in the DOGE/National Endowment for the Humanities (NEH) lawsuit to “claw back” the viral deposition videos they had posted to YouTube — videos showing DOGE operatives Justin Fox and Nate Cavanaugh stumbling through questions about how they used ChatGPT to decide which humanities grants to kill, and struggling mightily to define “DEI” despite it apparently being the entire basis for their work.

The government’s argument was that the videos had led to harassment and death threats against Fox and Cavanaugh — the same two who had no problem obliterating hundreds of millions in already approved grants with a simplistic ChatGPT prompt, but apparently couldn’t handle the public seeing them struggle to explain themselves under oath. The government argued the videos needed to come down. The judge initially agreed and ordered the plaintiffs to pull them. As we noted at the time, archivists had already uploaded copies to the Internet Archive and distributed them as torrents, because that’s how the internet works.

Well, now Judge McMahon has issued a full ruling on the government’s motion for a protective order, and has reversed course. The government’s motion is denied. The videos are now back up. There are hours and hours of utter nonsense for you to enjoy. Here are just a couple of the videos:

The ruling is worth reading in full, because McMahon manages to be critical of both sides while ultimately landing firmly against the government’s attempt to suppress the videos. She spends a good chunk of the opinion scolding the plaintiffs for what she clearly views as a procedural end-run — they sent the full deposition videos to chambers on a thumb drive without ever filing them on the docket or seeking permission to do so, which she sees as a transparent attempt to manufacture a “judicial documents” argument that would give the videos a presumption of public access.

McMahon doesn’t buy it:

When deciding a motion for summary judgment, the Court wants only those portions of a deposition on which a movant actually relies, and does not want to be burdened with irrelevant testimony merely because counsel chose to, or found it more convenient to, submit it. And because videos cannot be filed on the public docket without leave of court, there was no need for the rule to contain a specific reference to video transcriptions; the only way to get such materials on the docket (and so before the Court) was to make a motion, giving the Court the opportunity to decide whether the videos should be publicly docketed. This Plaintiffs did not do.

But if Plaintiffs wanted to know whether the Court’s rule applied to video-recorded depositions, they could easily have sought clarification – just as they could easily have filed a motion seeking leave to have the Clerk of Court accept the videos and place them on the public record. Again, they did not. At the hearing held on March 17, 2026, on Defendants’ present motion for a protective order, counsel for ACLS Plaintiffs, Daniel Jacobson, acknowledged the reason, stating “Frankly, your Honor, part of it was just the amount of time that it would have taken” to submit only the portions of the videos on which Plaintiffs intended to rely. Hr’g Tr., 15:6–7. In other words, “It would have been too much work.” That is not an acceptable excuse.

The Court is left with the firm impression that at least “part of” the reason counsel did not ask for clarification was because they wished to manufacture a “judicial documents” argument and did not wish to be told they could not do so. The Court declines to indulge that tactic.

Advertisement

Fair enough. But having knocked the plaintiffs for their procedural maneuver, the judge then turns to the actual question: has the government shown “good cause” under Rule 26(c) to justify a protective order keeping the videos off the internet? And the answer is a pretty resounding no. And that’s because public officials acting in their official capacities have significantly diminished privacy interests in their official conduct:

The Government’s motion fails for three independent reasons. First, the materials at issue concern the conduct of public officials acting in their official capacities, which substantially diminishes any cognizable privacy interest and weighs against restriction. Second, the Government has not made the particularized showing of a “clearly defined, specific and serious injury” required by Rule 26(c). Third, the Government has not demonstrated that the prospective relief it seeks would be effective in preventing the harms it identifies, particularly where those harms arise from the conduct of third-party actors beyond the control of the parties.

She cites Garrison v. Louisiana (the case that extended the “actual malice” standard from NY Times v. Sullivan) for the proposition that the public’s interest “necessarily includes anything which might touch on an official’s fitness for office,” and that “[f]ew personal attributes are more germane to fitness for office than dishonesty, malfeasance, or improper motivation.” Given that these depositions are literally about how government officials decided to terminate hundreds of millions of dollars in grants, that framing fits.

The judge also directly calls out the government’s arguments about harassment and reputational harm, and essentially says: that’s the cost of being a public official whose official conduct is being scrutinized. Suck it up, DOGE bros.

Reputational injury, public criticism, and even harsh commentary are not unexpected consequences of disclosing information about public conduct. They are foreseeable incidents of public scrutiny concerning government action. Where, as here, the material sought to be shielded by a protective order is testimony about the actions of government officials acting in their official capacities, embarrassment and reputational harm arising from the public’s reaction to official conduct is not the sort of harm against which Rule 26(c) protects. Public officials “accept certain necessary consequences” of involvement in public affairs, including “closer public scrutiny than might otherwise be the case.”

As for the death threats and harassment — which McMahon explicitly says she takes seriously and calls “deeply troubling” and “highly inappropriate” — she notes that there are actual laws against threats and cyberstalking, and that Rule 26(c) protective orders aren’t a substitute for law enforcement doing its job:

There are laws against threats and harassment; the Government and its witnesses have every right to ask law enforcement to take action against those who engage in such conduct, by enforcing federal prohibitions on interstate threats and cyberstalking, see, e.g., 18 U.S.C. §§ 875(c), 2261A, as well as comparable state laws. Rule 26(c) is not a substitute for those remedies.

And then there’s the practical reality McMahon acknowledges directly: it’s too damn late. The videos have already spread everywhere. A protective order aimed solely at the plaintiffs would accomplish approximately nothing.

At bottom, the Government has not shown that the relief it seeks is capable of addressing the harm it identifies. The videos have already been widely disseminated across multiple platforms, including YouTube, X, TikTok, Instagram, and Reddit, where they have been shared, reposted, and viewed by at least hundreds of thousands of users, resulting in near-instantaneous and effectively permanent global distribution. This is a predictable consequence of dissemination in the modern digital environment, where content can be copied, redistributed, and indefinitely preserved beyond the control of any single actor. Given this reality, a protective order directed solely at Plaintiffs would not meaningfully limit further dissemination or mitigate the Government’s asserted harms.

Separately, the plaintiffs asked for attorney’s fees, and McMahon denied that too, noting that she wasn’t going to “reward Plaintiffs for bypassing its procedures” even though the government’s motion ultimately failed. So everyone gets a little bit scolded here. But the bottom line is clear: you don’t get to send unqualified DOGE kids to nuke hundreds of millions in grants using a ChatGPT prompt, and then ask a court to hide the video of them trying to explain themselves under oath.

Releasing full deposition videos is certainly not the norm, but given that these are government officials who were making massively consequential decisions with a chatbot and no discernible expertise, the world is much better off with this kind of transparency — even if Justin and Nate had to face some people on the internet making fun of them for it.

Filed Under: depositions, doge, justin fox, nate cavanaugh, neh, public scrutiny

Companies: american council of learned societies, american historical association, authors guild