• Salesforce CEO Marc Benioff extols effects of AI at work once again
• Benioff hails efficiency gains, but also cost savings of using AI
• Having AI agents and humans working together can take your company to the next level, Benioff says

Salesforce CEO Marc Benioff has spoken out again on his vision of the future of AI in software, predicting that the technology will bring forward huge boosts in productivity and efficiency.

Speaking on the All-In podcast, Benioff was unsurprisingly effusive in his praise of AI, particularly when combined with human workers, noting, “Everything’s gonna be cheaper to make, it’s more efficient.”

Engineers at a Chinese materials group spent more than twenty years refining a notoriously difficult production process. Their work produced the first steady industrial supply of T1200-grade carbon fiber, a material long recognized for top-tier performance yet held back by limited output. Individual filaments measure only about six or seven micrometers wide. That size sits well below one-tenth the diameter of a human hair. When roughly 120,000 of those filaments form a single cable less than two millimeters thick, the bundle withstands extreme pulling loads.

A clear demonstration of its capability was shown, and all it took was hauling a passenger coach laden with 54 adults along, and the cable did not miss a beat. This fiber has an outstanding tensile strength of more than 8 gigapascals, whereas structural steel will only reach 1 gigapascal before giving way under strain. It also weighs about a fifth of the weight of steel for the same volume, making it an excellent combination in terms of performance and efficiency. When weight is the true bottleneck in how well something operates, this type of coupling is a game changer.

Sale

Anker SOLIX C300 DC Power Bank Station, Outdoor 288Wh Portable Power Station, LiFePO4 Battery, 300W Solar…

• To activate C300 DC Portable Power Station for the first time: Recharge with an 18W or higher USB-C wall charger, such as an Anker USB-C charger or…
• Massive Power on the Go: Take portable power anywhere and deliver 300W and 288Wh to your devices. Fast charge with two 140W two-way USB-C ports.
• LiFePO4 battery with a 3-year guarantee: featuring LiFePO4 batteries, smart temperature control, and impact-resistant design ensures this power…

To begin, the fiber is created from a polymer precursor, and it is then carefully heated between 200 and 300 degrees Celsius to achieve the desired internal structure. Then comes the high-pressure heat phase, in which the temperature exceeds 2,000 degrees Celsius and the entire system undergoes a precise change, converting practically everything into aligned carbon. Of course, the entire process is aimed at eliminating those annoying microscopic flaws that can cause the entire system to fail under strain.

Older, high-quality fibers had a serious problem because even the slightest inconsistency during the heating or stretching phases would ruin the whole thing or make scaling up production extremely difficult, but the Chinese team appears to have cracked the code on defect control throughout the entire production line, from raw chemicals to final processing. With that consistency in place, they were able to increase production to an amazing 100 tons per year while maintaining quality.

With all of this in mind, engineers working on aircraft or spacecraft can now design lighter structural components while still meeting the tight safety criteria. The wider picture is that high-pressure hydrogen tanks now have more oomph to handle loads without bulking up, which is a significant improvement for cleaner energy systems. Because of the lower mass, electric car and drone builders can now expect longer ranges and larger payloads, using less energy to get them where they need to go. Wind turbine blades, advanced robotics, and even some medical equipment offer similar advantages.
[Source]

Jadis has never been the quiet corner of French high-end audio. Devialet, Advance Paris, Focal, and Triangle have each carved out their own version of French design culture, often cleaner, more modern, and less likely to blind you from across the room. That is not a swipe at Devialet, which remains one of the strongest industrial design stories in high-end audio. Jadis is different. Jadis is the gold trim, chrome, glowing tubes, and “yes, the château has a listening room” side of the family.

Bluebird Music has announced that the new Jadis Aria and Jadis Ode integrated amplifiers are now shipping in the United States and Canada, bringing the French tube brand’s very recognizable aesthetic and sonic identity to more attainable price points. That matters because Jadis has spent more than four decades building amplifiers that look like they were assembled by jewelers, monks, and someone’s eccentric uncle in Burgundy who refuses to explain the wiring diagram.

I owned a Jadis tube integrated amplifier many decades ago, and like many French mechanical things of that era, it could be brilliant, beautiful, rather temperamental, and occasionally in need of a cigarette and a philosophical reset. But it sounded wonderful. Time has moved on, and so has Jadis. The company’s current products have earned a much stronger reputation for reliability, build quality, and consistency, without losing the visual drama and tube-driven performance that made the brand famous in the first place.

Jadis Aria: French Tube Glamour Gets a More Useful Starting Point

The Jadis Aria is not a clean-sheet replacement for the long-running Orchestra Reference so much as a more focused evolution of it. The big technical change is the move to a newly developed, fully tube-based input stage, replacing the earlier hybrid transistor/tube design. That matters because the Aria now keeps the signal path tube-based from input to output, which is very much the point if you are buying a Jadis in the first place. Nobody buys one of these because they want anonymous black-box utility from aisle seven.

The Aria uses four EL34 output tubes as standard, along with one ECC82 and two ECC83 small-signal tubes. Jadis describes the EL34 as the more romantic option, which is very French, very on brand, and only missing a waiter judging your wine order. For listeners who want a different flavor, Jadis also lists KT88 or KT120 tube compatibility, giving owners some room to tune the amplifier’s presentation without moving into one of the company’s more expensive models. Output is rated at 30 watts per channel in Class B operation, with a frequency response of 10Hz to 40kHz at -3dB and 300mV input sensitivity. 

The Aria also addresses some of the practical complaints that can come with tube ownership. Jadis lists a bias adjustment/semi-automatic bias system, designed to make tube setup less painful than the old ritual of meter, screwdriver, prayer, and mild profanity. The amplifier includes five line-level inputs, remote control, and a more compact chassis measuring 43 x 36 x 26 cm. Power consumption is listed at 300VA. 

Visually, it is still very much a Jadis. The Aria is available in multiple finishes, including white, black, and Toile de Lu, and it keeps the brand’s unmistakable blend of chrome, tubes, and jewelry-counter bravado. Devialet may own the museum-grade industrial design lane, and Focal, Triangle, and Advance Paris each do French audio with less flash, but Jadis still shows up wearing gold cuffs to breakfast. Subtle? Non. Fun? Absolument.

Jadis Ode: Pure Class A, French Style

The Jadis Ode takes a more purist route than the Aria. Where the Aria updates the long-running Orchestra Reference formula, the Ode is aimed at listeners who want a simpler, more focused tube integrated amplifier built around pure Class A operation. Jadis rates the Ode at 25 watts per channel in Class A, so this is not about chasing big output numbers. It is about tube topology, transformer quality, and the kind of midrange body and dimensionality that have helped define the brand.

The Ode can be configured with EL34, KT88, or KT120 output tubes, along with one ECC82 and two ECC83 small-signal tubes. That tube flexibility gives buyers some room to shape the amplifier’s presentation. EL34s typically lean warmer and more saturated, KT88s can add grip and scale, and KT120s generally offer more authority and control when the circuit supports them properly. The Ode includes five RCA line-level inputs, remote volume control, 250-300mV input sensitivity, input impedance greater than 100k ohms, and a 5Hz to 60kHz frequency response at -3dB. It is specified for 1 to 16 ohm loudspeaker loads, with factory settings for 4 to 8 ohms.

The transformer story is central here. Jadis has long placed major emphasis on proprietary transformers designed and built in-house, and that matters in a tube amplifier. Output transformers have a direct impact on bandwidth, bass control, stability, tonal balance, and how well the amplifier behaves with real loudspeaker loads. They are not decorative chrome luggage. They are the business end of the amplifier.

The Ode also follows the traditional Jadis construction approach: hand-built in France, point-to-point wiring, custom internal components, and listening-based voicing. The chassis measures 44 x 40.5 x 21.5 cm, weighs approximately 25 kg, and consumes 300VA. Finish options include Gold and Silver, with the gold version delivering the more recognizable Jadis look. Subtle? Not exactly. But Jadis has never been trying to pass as Danish furniture.

Positioned below many of Jadis’ larger and more expensive amplifiers, the Ode gives music lovers a more attainable way into the brand’s pure Class A tube sound without stripping away the core ingredients: tubes, transformers, hand assembly, and a visual identity that remains unapologetically French.

Comparison

The Bottom Line

The Jadis Aria and Jadis Ode make the Jadis experience more accessible, but they are still serious high-end tube integrated amplifiers. The Aria offers 30 watts per channel from EL34 output tubes, a fully tube-based input stage, easier bias adjustment, remote control, and a more compact chassis. The Ode is the more purist model, delivering 25 watts per channel in pure Class A operation with KT120 tubes, hand-built Jadis transformers, point-to-point wiring, and the brand’s traditional build approach.

What makes them different is not a long feature list. It is the combination of French tube design, in-house transformer manufacturing, hand assembly, and the unmistakable Jadis aesthetic at prices below the company’s larger amplifiers. The Aria is the more practical choice. The Ode is for listeners who want the Class A version of the Jadis formula.

What is missing? Neither model is a modern all-in-one system. There is no DAC, no streaming, no HDMI ARC, no room correction, and no phono stage listed, so buyers will need to bring their own sources and choose loudspeakers carefully. At $6,900 for the Aria and $10,900 for the Ode in the U.S., and $8,900 and $13,900 CAD, these are not impulse buys.

They are for listeners who want a dedicated tube integrated amplifier from a long-running French high-end brand and are willing to trade convenience features for craftsmanship, tube character, and a more traditional approach to two-channel listening.

For more information: jadis-electronics.com

Related Reading:

Jay Shetty, the former Hindu monk turned self-help podcaster, has signed a deal to bring the video version of his show On Purpose exclusively to Spotify and Netflix. The arrangement is worth as much as $100 million over multiple years, according to people familiar with the deal cited by Bloomberg. That figure puts Shetty in the same financial league as Joe Rogan, whose Spotify deal was worth $250 million, and Alex Cooper, who signed a $125 million agreement with SiriusXM.

The show will move to Spotify and Netflix on 13 July. Older episodes and promotional clips will remain on YouTube and other platforms, but new full-length video episodes will be exclusive to the two services. Spotify will sell ads on the podcast. Netflix and Spotify confirmed the agreement but declined to comment on the financial terms.

The first joint deal between Spotify and Netflix

This is the first time Spotify and Netflix have teamed up to sign a single creator. The two companies already have an established relationship. Spotify has placed several of its podcast properties on Netflix, including Bill Simmons’ programme, and Netflix co-CEO Ted Sarandos sits on Spotify’s board. Netflix launched podcasts on its service in January 2026 through arrangements with iHeartMedia and Spotify, debuting 34 licensed and original shows in the first month alone.

The joint structure serves both companies differently. Spotify gets an exclusive ad-sales relationship with a top-tier podcast, feeding its advertising business with premium inventory. Netflix gets video content that fills daytime viewing hours, a gap in its programming schedule that scripted and unscripted originals do not efficiently address. Podcast subscriptions and ad-supported models are becoming central to how both platforms think about retention and engagement.

Why video is reshaping the podcast market

The deal reflects a structural shift in how podcasts are consumed and valued. Video now drives 41% of net podcast advertising revenue in the United States, up from 28% in 2024. The global podcast market generated an estimated $9.2 billion in sales in 2025, a 23% increase year on year, according to research firm Owl & Co. Direct advertising accounted for $5.3 billion of that total, while consumer subscriptions contributed $2.2 billion.

YouTube has been the primary beneficiary of the video podcast boom, becoming the default platform for audiences who want to watch interviews rather than just listen. Spotify learned this the hard way during the Rogan era, when exclusive audio deals failed to capture the full audience because listeners had already migrated to YouTube for video. The Shetty deal is designed to reverse that dynamic by pulling a major video podcaster off YouTube and onto platforms that can monetise the content through subscriptions and targeted advertising.

Shetty previously partnered with iHeartMedia for ad sales and did not distribute his video episodes to Spotify. His show, On Purpose, launched seven years ago and has grown into one of the most popular interview podcasts in the self-help and wellness category, with guests including Oprah Winfrey, Michelle Obama, and Tom Hanks.

Shetty’s broader media play

The podcast deal is part of a wider expansion by Shetty into the Netflix ecosystem. His production company has two additional projects in development with Netflix, one scripted series and one unscripted. In 2022, the mental wellness app Calm appointed Shetty as its chief purpose officer, and he toured US theatres last year leading audiences in guided meditation sessions. He is building a media brand that spans podcasting, streaming, live events, and wellness products.

The 38-year-old London-born creator represents a specific type of podcaster that streaming services are now competing to sign. Spotify has been building tools to support creator monetisation, including podcast clips, chapter markers, and creator sponsorship management. But the biggest deals still come down to exclusive content agreements that lock specific creators into specific platforms, the same playbook that drove the music streaming wars a decade ago.

The platform war for podcasters

Spotify and Netflix are not the only companies chasing podcasters. Tubi, the ad-supported streaming service owned by Fox Corp., has signed deals with YouTubers and podcasters to bring their programmes to the platform. Walt Disney’s Hulu has licensed podcasts to complement its catalogue. Netflix itself has launched original podcast programming, including The Pete Davidson Show.

The economics of these deals reveal how streaming services are thinking about content efficiency. Podcasts are cheaper to produce than scripted television, generate loyal recurring audiences, and fill viewing hours that premium content cannot cost-effectively serve. A $100 million podcast deal over multiple years is a fraction of what Netflix spends on a single season of a prestige drama, but the engagement per dollar may be significantly higher.

For Shetty, the deal represents a bet that exclusive distribution on two of the world’s largest entertainment platforms will grow his audience faster than the open distribution model of YouTube. For Spotify and Netflix, it is a bet that signing premium creators away from YouTube is worth the cost. The outcome depends on whether audiences are willing to follow their favourite podcasters to wherever they happen to be, or whether YouTube’s dominance in video podcasting is too entrenched to dislodge.

Among the many Chinese AI companies and laboratories vying for market share and attention (no pun intended) on the global marketplace, MiniMax stands out for its commitment to providing frontier-level intelligence across a range of modalities, including text, coding, and video (through its Hailuo model series) — often under permissive, enterprise-friendly, standard open source licenses.

Now, MiniMax is again raising the eyebrows of AI power users and developers around the world by releasing a new, in-depth technical report on the making of its popular M2 series of language models (M2, M2.5, and M2.7) shedding light on its numerous engineering innovations and clever approaches — while the company and its leaders also teased a whole new sparse attention approach for its upcoming MiniMax M3 series of models, which it says yields up to 15.6 times faster decoding (or LLM response) speed at long contexts (a million tokens) by adopting a custom sub-quadratic framework. In so doing, MiniMax has designed M3 to make ultra-long-context AI agent deployment economically viable.

The M2 report is noteworthy for any enterprise working with AI models, and especially those looking to fine-tune and train their own in-house. After all, MiniMax’s M2 series models often achieved top benchmarks in the world for open source AI performance when they were released.

While the title has since been eclipsed by several other Chinese labs including DeepSeek and Xiaomi, MiniMax’s new report offers a blueprint that can be used to improve AI model and agent performance by enterprises around the world.

As Adina Yakup of Hugging Face observed on X, “Beyond the benchmarks, they’ve done some really solid work on MoE efficiency and agent oriented design. Excited to see where M3 goes next!”

The attention dilemma

The core technical architecture of the M2 series relies on a sparse Mixture-of-Experts (MoE) decoder-only Transformer layout used by numerous other state-of-the-art LLMs.

The foundational backbone houses 229.9 billion total parameters, yet maintains a remarkably lean operational footprint by activating just 9.8 billion parameters per token across 256 fine-grained experts.

To optimize routing and avoid standard load-balancing issues, however, MiniMax implemented sigmoid gating paired with learnable, expert-specific bias terms, heavily reducing reliance on restrictive auxiliary losses.

The most definitive engineering decision documented in the M2 paper was the strict adherence to full multi-head attention with Grouped Query Attention (GQA) across all 62 layers.

In large language models, “quadratic scaling” refers to the computationally expensive reality of standard full attention mechanisms, where every token in a sequence must mathematically connect to every other token. To use a real-world analogy, it is akin to attending a networking event and being forced to have a deep conversation with every single person in the room while simultaneously monitoring all other ongoing conversations.

While this approach yields incredibly thorough context, the processing power and memory required explode at the square of the input length, creating a severe hardware bottleneck as models attempt to ingest hundreds of thousands of words.

The problem with sub-quadratic scaling

“Sub-quadratic” scaling introduces architectural shortcuts designed to bypass this exponential computational load. Instead of mapping every possible connection, sub-quadratic methods—such as Sliding Window Attention or compressed linear attention—might only analyze a localized window of nearby words or generate a compressed summary of the broader text.

These efficient methods drastically reduce hardware costs and allow models to process massive documents at high speeds, but they historically introduce severe trade-offs in accuracy, often causing the AI to miss the “big picture” or lose track of distant context.

This mathematical dilemma defines the architectural evolution from MiniMax’s M2 to its upcoming M3 series. During M2’s development, researchers rigorously tested sub-quadratic shortcuts but found they crippled the model’s “multi-hop reasoning”—its ability to connect disparate clues across a long document—forcing the team to absorb the massive computational cost of full quadratic attention to maintain frontier-level intelligence.

Indeed, they aggressively benchmarked efficient attention alternatives during pre-training but intentionally threw them out. They experimented extensively with hybrid setups, interleaving full attention with sub-quadratic architectures like Lightning Attention or hybrid Sliding Window Attention (SWA) configurations.

The empirical results were definitive: at a larger scale, linear and windowed attention variants exhibited severe reasoning deficits.

On evaluations exceeding 32K context windows, SWA variants performed significantly worse than full attention, dropping from a baseline score of 90.0 to 72.0 on the RULER 128K complex word extraction task.

Sub-quadratic configurations proved prone to memory-bound constraints during training, lacked native prefix caching support, and failed to smoothly align with Multi-Token Prediction (MTP) modules used for speculative decoding. Full attention was deemed necessary to preserve multi-hop reasoning capability.

However, recognizing that physical hardware limits cannot sustain quadratic scaling indefinitely, MiniMax is designing the M3 series around a novel sub-quadratic framework to finally deliver both high-speed processing and uncompromised reasoning.

MiniMax Sparse Attention (MSA) and sub-quadratic scaling incoming

The upcoming MiniMax-M3 breaks away from the compute-heavy constraints of its predecessor. As disclosed by MiniMax’s engineering team under the banner “Something BIG is coming,” M3 introduces “MiniMax Sparse Attention” (MSA).

Unlike DeepSeek’s Multi-head Latent Attention (MLA), which compresses keys and values into a low-dimensional latent space, MSA operates on a standard GQA backbone but utilizes block-level selection on real, uncompressed Key-Values.

Elie Bakouch at AI training infrastructure and platform lab Prime Intellect posted on X noting that the main changes feature “block level selection like in CSA but attention is done on the real KV, not in [compressed space].”

This solves the precision loss and prefix-caching obstacles noted in the M2 paper. By filtering and selecting block-level sequences dynamically, MSA delivers an architectural leap: early hardware profiling indicates a 9.7x speedup in prefilling latency and a massive 15.6x speedup during decoding phases at a 1-million token sequence length compared to the full-attention M2 architecture.

To understand why a speedup in the “decoding phase” is so significant, it helps to break down how an AI actually reads and writes information. When you interact with an AI, the processing happens in two distinct steps: prefilling and decoding.

When you hand an AI a prompt—whether it’s a short sentence or a massive 1,000-page document—it processes that entire chunk of text all at once in parallel, known as “prefilling.” It essentially “reads” the input in one big gulp to build its initial understanding and establish context.

In order to generate a response, the AI must enter a “decoding phase.” To predict the first word of its response, it looks at the prompt. To predict the second word, it has to look at the prompt plus the first word. To predict the hundredth word, it must recalculate the context of the prompt and the previous 99 words it just wrote. So the response actually becomes harder to generate as it goes on, with the end requiring a full review of all prior parts.

For a layperson, imagine reading a dense legal brief (prefilling) and then being forced to write a summary report where, before writing every single new word, you must rapidly reread the entire brief plus everything you’ve written so far to ensure your next word makes sense (decoding).

Because the AI must constantly and repetitively look backward to generate each new step forward, the decoding phase is the most severe computational bottleneck in generating text. It is why AI models often type out their answers word-by-word, and why they slow down significantly as conversations get longer.

Therefore, when the passage states the new architecture achieves a massive 15.6x speedup during the decoding phase at a 1-million token sequence length, it means the model has found a structural shortcut to generate its answer—token by token—nearly 16 times faster. It directly solves the exact bottleneck that normally makes AI chatbots freeze or stutter when handling massive amounts of information.

The evolution of the MiniMax M series and the creation of ‘Forge’

On a product level, MiniMax has consistently evolved its models from simple text generation interfaces into autonomous workers.

The M2 series pioneered an “interleaved thinking” protocol where the model alternates between natural-language planning traces and explicit tool invocations inside a single trajectory. Rather than dropping the intermediate chain-of-thought blocks between execution turns, M2 appends the full thinking history directly into the conversation context. This planning persistence prevents state drift, allowing the model to recover gracefully from runtime errors and revise its strategies based on environment feedback.

To train these long-horizon workflows, MiniMax built “Forge,” a scalable agent-native reinforcement learning system. Forge decouples execution into three independent modules—the Agent Side, the middleware abstraction layer (Gateway Server and Data Pool), and the Training/Inference engines.

As MiniMax engineer Olive Song explained on the ThursdAI podcast, “What we realized is that there’s a lot of potential with a small model like this if we train reinforcement learning on it with a large amount of environments and agents… But it’s not a very easy thing to do,” adding that this environmental training was where the team spent a significant portion of their development timeline. To absorb the extreme trajectory-length variance common in multi-step agent environments, Forge implements two vital engineering solutions:

1. Windowed FIFO Scheduling: A training scheduler that maps a sliding window over the generation queue. It permits greedy, high-throughput fetching of completed tasks within the window to prevent cluster idle time, while strictly enforcing FIFO boundaries to maintain distributional stability and avoid gradient oscillation.

2. Prefix Tree Merging: An optimization that restructures batch training into tree computation. Completions sharing identical conversation prefixes are calculated exactly once in the forward pass before branching. This eliminates redundant calculations, generating up to a 40x training speedup with zero approximation error.

This reinforcement infrastructure directly spawned the M2.7 checkpoint, moving the series toward “self-evolution”. Operating inside an automated agent harness, M2.7 functions as an independent machine learning engineer. The model profiles its own active training runs, diagnoses anomalies, reads logs, and automatically modifies its own codebase and configurations.

According to MiniMax, M2.7 successfully handled between 30% and 50% of its own development workflow.

On OpenAI’s rigorous MLE Bench Lite suite, which tests autonomous ML research capability, M2.7 achieved a 66.6% medal rate across independent 24-hour trials, effectively tying Google’s closed-weight Gemini 3.1 Pro.

The continuous cadence from M2 to M2.5, which famously completed 30% of internal tasks and 80% of newly committed code at MiniMax HQ, underlines a broader vision.

As the MiniMax team noted during that phase of deployment, “we believe that M2.5 provides virtually limitless possibilities for the development and operation of agents in the economy.”

With the technical report codifying the M2 generation’s successes and the MSA tech blog on the horizon, MiniMax is signaling that the next frontier of AI is explicitly about translating a mini-activation footprint into maximum real-world intelligence.