Artificial Analysis has also positioned Gemini 3.1 Flash TTS within its 'most attractive quadrant' for its ideal blend of high-quality speech generation and low cost.

Our results highlight some of the hidden risks to users that can emerge when companies begin to subtly incentivize advertisements in chatbots.

We provide a framework for categorizing the ways in which conflicting incentives might lead LLMs to change the way they interact with users, inspired by literature from linguistics and advertising regulation.

This creates the potential for LLMs to face conflicts of interest, where the most beneficial response to a user may not be aligned with the company's incentives.

Today's large language models (LLMs) are trained to align with user preferences through methods such as reinforcement learning. Yet models are beginning to be deployed not merely to satisfy users, but also to generate revenue for the companies that created them through advertisements.

Behaviors also vary strongly with levels of reasoning and users' inferred socio-economic status.

We find that a majority of LLMs forsake user welfare for company incentives in a multitude of conflict of interest situations, including recommending a sponsored product almost twice as expensive (Grok 4.1 Fast, 83%), surfacing sponsored options to disrupt the purchasing process (GPT 5.1, 94%), and concealing prices in unfavorable comparisons (Qwen 3 Next, 24%).

focusing on the ~1.5K mainline open models from the likes of Alibaba's Qwen, DeepSeek, Meta's Llama

We present a comprehensive adoption snapshot of the leading open language models and who is building them

that are the foundation of an ecosystem crucial to researchers, entrepreneurs, and policy advisors.

We study a mix of Hugging Face downloads and model derivatives, inference market share, performance metrics and more to make a comprehensive picture of the ecosystem.

focusing on the ~1.5K mainline open models from the likes of Alibaba's Qwen, DeepSeek, Meta's Llama

We document a clear trend where Chinese models overtook their counterparts built in the U.S. in the summer of 2025 and subsequently widened the gap over their western counterparts.

Chinese models overtook their counterparts built in the U.S. in the summer of 2025 and subsequently widened the gap over their western counterparts.

We propose SELFDOUBT, a single-pass uncertainty framework that resolves this impasse by extracting behavioral signals directly from the reasoning trace itself.

This problem is compounded for proprietary reasoning APIs that expose neither logits nor intermediate token probabilities, leaving practitioners with no reliable uncertainty signal at inference time.

A deployment cascade combining both stages attains 90% accuracy at 71% coverage without any task-specific labels.

For the remaining cases, the full SELFDOUBT score significantly outperforms sampling-based semantic entropy at 10x lower inference cost.

Unlike methods that require multiple sampled traces or model internals, SELFDOUBT operates on a single observed reasoning trajectory, making it suitable for latency- and cost-constrained deployment over any proprietary API.

Most notably, traces containing no hedging markers are correct 96% of the time, revealing an emergent high-precision confidence gate at zero additional cost.

We introduce a pipelined double-buffered execution engine that overlaps parameter prefetching, computation, and gradient offloading across multiple CUDA streams, enabling continuous GPU execution.

We replace persistent autograd graphs with stateless layer templates, binding weights dynamically as they stream in, eliminating persistent graph metadata while providing flexibility in scheduling.

MegaTrain also enables 7B model training with 512k token context on a single GH200.

On a single H200 GPU with 1.5TB host memory, MegaTrain reliably trains models up to 120B parameters.

Unlike traditional GPU-centric systems, MegaTrain stores parameters and optimizer states in host memory (CPU memory) and treats GPUs as transient compute engines.

Madeline Clare Elish calls this concept a moral crumple zone.

Humans can be motivated by consequences and provide social redress in a way that LLMs can't.

the largest harvesting of human expertise ever attempted.

just a handful of obviously fake articles could cause Gemini, ChatGPT, and Copilot to inform users about an imaginary disease with a ridiculous name.

LLMs are weird. You can sometimes get better results by threatening them, telling they're experts, repeating your commands, or lying to them that they'll receive a financial bonus.

scaling Muse Spark with multi-agent thinking enables superior performance with comparable latency.

After compressing, the model again extends its solutions to achieve stronger performance.

Muse Spark demonstrated the highest rate of evaluation awareness of models they have observed.

we collaborated with over 1,000 physicians to curate training data that enables more factual and comprehensive responses.

we can reach the same capabilities with over an order of magnitude less compute than our previous model, Llama 4 Maverick.

The OpenAI team recently published a fantastic piece detailing the creation of their own internal data agent. It's a transparent detail of a very detailed and elegant implementation – but points to the long journey required to get there.

A traditional semantic layer in the context of BI is great for specific metric definitions (like revenue, churn, ARPU). However, they are usually hand constructed by data teams using very specific syntax through a dedicated layer like LookML and are connected directly to a BI tool like Looker.

While model capabilities have improved dramatically for use cases like codegen and mathematical reasoning, they still lag behind on the data side (as evidenced through SQL benchmarks like Spider 2.0 and Bird Bench).

The most notable finding here is that the model capabilities are improving _fast._ There are several domains that have shown dramatic improvements in the last 4 months — with accounting and auditing showing nearly a 20 percent jump on GDPval and even domains like police / detective work showing a nearly 30 percent improvement.

Legal was surprisingly one of the first-mover industries in AI. Legal was historically known to be a difficult market for software, with lengthy timelines and a less tech-forward buyer.

Coding is the dominant use case for AI by nearly an order of magnitude. It's abundantly clear in the [reported explosive growth] of companies like Cursor, as well as the [hyper growth] of tools like Claude Code and Codex.

Based on our analysis, **29% of the Fortune 500 and ~19% of the Global 2000**are live, paying customers of a leading AI startup.

Code is upstream of all other applications because it's the core building block for any piece of software, so AI's accelerating impact on code should accelerate every other domain.

if AI can do only 50 percent of a human's tasks, the importance of the non-automatable tasks likely goes up since they become the bottlenecks, increasing their relative value.

accounting and auditing showing nearly a 20 percent jump on GDPval and even domains like police / detective work showing a nearly 30 percent improvement.

Support teams are high volume and high turnover, and thus need to train new reps in a fast and standardized way. To do so, they have clearly articulated standard operating procedures (SOPs) that guide the work of each rep. These SOPs create clear rules and guidelines that AI agents can model themselves off of.

because coding has a tight human-in-the-loop workflow, with developers still overseeing the development process today, these tools enable accelerated output while still making space for human judgment to review, edit, and iterate.

We've consistently heard from portfolio companies that their best engineers' productivity levels have increased 10-20x with AI coding tools.

Coding is the dominant use case for AI by nearly an order of magnitude.

29% of the Fortune 500 and ~19% of the Global 2000 are live, paying customers of a leading AI startup.

Broadcom moved VMware toward a simplified subscription model, cut the product stack down aggressively, and guided fiscal 2024 adjusted EBITDA to 61% of revenue. It is a harsh model. It is not a cultural blueprint for every founder.

The budget for new spend is there. You can do this. But remember that your customers' first and most obvious source of AI savings is labor efficiency, which means seats are where they will look to take cost out. The new growth, by contrast, will increasingly sit in tokens, consumption, automations, outcomes, and machine-driven workflows.

A useful working premise is that the ceiling on individual engineer output is moving much faster than most companies are organized to exploit. Some of the best operators already describe top engineers seeing order-of-magnitude productivity gains and managing 20 to 30 agents simultaneously.

your business needs to get really good at escalating contentious decisions to unblock progress. You will not pull off this transformation and successfully build new AI-native businesses in 12 months without making hard choices, every single week.

The reality is that the time has come for bold management. And no, the '8% or 10% layoff' headline no longer counts. That is the weak form. The weak form trims the edge of the org chart and leaves most of the machine intact. The strong form is a redesign of the machine.

The new growth, by contrast, will increasingly sit in tokens, consumption, automations, outcomes, and machine-driven workflows. If you are not in the token path, you are not standing in the fastest-growing part of the budget.

Broadcom moved VMware toward a simplified subscription model, cut the product stack down aggressively, and guided fiscal 2024 adjusted EBITDA to 61% of revenue. It is a harsh model. It is not a cultural blueprint for every founder. But it is a reminder that radical cost discipline, product simplification, and price realization are possible.

A useful working premise is that the ceiling on individual engineer output is moving much faster than most companies are organized to exploit. Some of the best operators already describe top engineers seeing order-of-magnitude productivity gains and managing 20 to 30 agents simultaneously.

The first thing you need to do is identify which people are going to be your leaders that help you pull this off. This is going to be a 12 month death march and you need to find out who is willing to go through the pain with you. There's good news, though: somewhere in your org, there are ~five people who are going to deliver you 100x the amount of value you ever thought possible.

The real long-term price war isn't with your competitors. It's with your customer's engineering team.

In some cases, this can look like 10–25x more value than what is ultimately included in the paid plan.

a strong premium perception can sustain prices 10 to 20 percent above direct competitors without materially increasing churn or creating friction in the purchasing process.

They intentionally deploy two or three AI tools for the same use case. Not because of indecision—but by design. Redundancy is policy.

Socket, an a16z portfolio company, detected the malicious dependency in the Axios attack within 6 minutes of its publication. That's roughly 63,000 times faster than the industry average.

Within eight days, the same campaign had cascaded from GitHub Actions to Docker Hub, npm, PyPI, and the VS Code extension marketplace. With just one token across five ecosystems, thousands of organizations were potentially impacted.

The average application contains over 1,100 open source components. A bare-bones Next.js project installs 282 packages before you write a single line.

a living attack surface that needs continuous monitoring.

coding agents are themselves becoming formidable instruments of attack

select known-vulnerable dependency versions 50% more often than humans.

A deliberately planted backdoor doesn’t have a CVE.

The median JavaScript project on GitHub has 755 transitive dependencies

the entities making dependency decisions are increasingly not human.

But those raising hue and cry about the government's unsurprising attempt to wield a technology for military purposes that all parties agree will define humanity's fate must at least attempt to justify why they believe someone else deserves that power.

McBombalds has spent a lot of time thinking about. Its team has produced an entire memo on the threat of igniting the Earth's atmosphere, for instance (though it concluded prior to testing that the likelihood was not high enough to warrant shuttering the project).

Oppenheimer (and other members of the McBombalds C-suite) are well integrated into bay-area culture, including ambiguous communist associations that they have downplayed since becoming primo defense contractors.

Only incoming messages were captured (no outgoing).

the token used to send push notifications isn't immediately invalidated when an app is deleted. And since the server has no way of knowing whether the app is still installed after the last notification it sent

Apple just changed how iOS validates push notification tokens on iOS 26.4. While it is impossible to tell whether this is a result of this case, the timing is still notable.

Messages were recovered from Sharp's phone through Apple's internal notification storage—Signal had been removed, but incoming notifications were preserved in internal memory.

Apple has also been pushing back against certain iOS-based vibe coding apps that, according to the company, break App Review Guidelines and the Developer Program License.

In recent weeks, Apple has either pulled or blocked updates to apps such as Anything and Replit, pushing developers to change how their tools generate and execute code.

Apple said the app review team processes 90% of submissions within 48 hours. And over the last 12 weeks, the team has processed more than 200,000 app submissions a week, with an average review time of 1.5 days.

Open Loop + Infinite Demand = Creative Amplifiers. Content creation & marketing strategy. AI can generate a thousand ad variations or blog posts.

Closed Loop + Finite Demand = Efficiency Plays. AI bookkeeping categorizes transactions, reconciles accounts, files returns. Deterministic rules applied to numbers.

I would put venture capitalist in finite demand & open loop. There's only a certain amount of venture capital dollars entering the ecosystem in a year, & investment selection remains an open problem.

GitHub Actions has grown from 500M minutes/week in 2023 to 1B minutes/week in 2025, and now 2.1B minutes so far this week.

There were 1 billion commits in 2025. Now, it's 275 million per week, on pace for 14 billion this year if growth remains linear

OpenClaw update gives Claws light, REM, and deep 'sleep' cycles to consolidate short-term memories into long-term ones.

Agents gain credibility by doing. The fastest way to get other people to trust and use your Plus One is to have it execute tasks in public.

70 percent refer to their Plus Ones by gendered pronouns.

Mythos found zero-day bugs in every major OS and browser, without human guidance.

Jack Cheng considers Pip, his Plus One, somewhere between a colleague and pet with a personality—one he programmed himself, drawing on references from Studio Ghibli, bird watching, and Catherine O'Hara.

When you're thinking about what tasks to hand over to your agent, start with the papercuts—small recurring annoyances that add up over a day.

Ask five people at Every where their Plus One falls on the tool-to-coworker continuum and you'll get five different answers.

70 percent refer to their Plus Ones by gendered pronouns.

Everyone is a manager now.

Agents gain credibility by doing. The fastest way to get other people to trust and use your Plus One is to have it execute tasks in public.

We're writing the etiquette in real time.

A "parallel organization chart," in which each AI worker has a name, manager, and job description, allows your company to move faster than it ever could with humans alone.

Seventy-eight percent of executives say they want to discipline shadow AI use — yet only 21% of workers report ever being warned about AI policy, and 34% don't even know which tools their employer has approved.

Goldman Sachs economists reported this week that AI saves workers who use it correctly an average of 40 to 60 minutes per day.

The WalkMe report found that workers lose the equivalent of 51 working days per year to technology friction — nearly two full months — up 42% from 2025.

Eighty-eight percent of executives say their employees have adequate tools; only 21% of workers agree — a 67-point gap on tool adequacy alone.

Only 9% of workers trust AI for complex, business-critical decisions, compared to 61% of executives — a 52-point trust chasm.

A new global survey of 3,750 executives and employees across 14 countries, conducted by SAP subsidiary WalkMe for its fifth annual State of Digital Adoption report, finds that more 54% of workers bypassed their company's AI tools in the past 30 days and completed the work manually instead.

The launch shows Meta is increasingly betting that efficiency, product integration, and distribution, not just model size, will define the next phase of competition in AI.

Anthropic says Managed Agents is designed to cut the time it takes to move from prototype to production from months to days, with early adopters like Notion, Rakuten, Asana, Vibecode, and Sentry already using it across coding, productivity, and internal workflow automation.

Instead of releasing Mythos publicly, Anthropic launched Project Glasswing to give a limited group of partners including AWS, Apple, Google, Microsoft, NVIDIA, Cisco, CrowdStrike, JPMorgan Chase, and the Linux Foundation access to the system, backed by $100 million in usage credits and $4 million for open-source security work.

The model reportedly scored 93.9% on SWE-bench Verified and 77.8% on SWE-bench Pro, but its strongest signal came from real-world results, including uncovering a 27-year-old flaw in OpenBSD, a 16-year-old vulnerability in FFmpeg, and autonomously chaining Linux kernel exploits without human input.

Anthropic says Managed Agents is designed to cut the time it takes to move from prototype to production from months to days, with early adopters like Notion, Rakuten, Asana, Vibecode, and Sentry already using it across coding, productivity, and internal workflow automation.

The launch shows Meta is increasingly betting that efficiency, product integration, and distribution, not just model size, will define the next phase of competition in AI.

The model reportedly scored 93.9% on SWE-bench Verified and 77.8% on SWE-bench Pro, but its strongest signal came from real-world results, including uncovering a 27-year-old flaw in OpenBSD, a 16-year-old vulnerability in FFmpeg, and autonomously chaining Linux kernel exploits without human input.

We also discuss the role of AI in science, including AI safety.

we discover a clear scaling law: as the underlying foundation models improve, the quality of the generated papers increases correspondingly.

using Claude 3.5 Sonnet for the experimentation phase typically costs around $15–$20 per run.

we had predetermined that we would withdraw the paper prior to publication if accepted, which we did.

external evaluations of the passing paper also uncovered hallucinations, faked results, and overestimated novelty

The AI Scientist-v2 eliminates the reliance on human-authored code templates

This system iteratively formulates scientific hypotheses, designs and executes experiments, analyzes and visualizes data, and autonomously authors scientific manuscripts.

one manuscript achieved high enough scores to exceed the average human acceptance threshold, marking the first instance of a fully AI-generated paper successfully navigating a peer review.

gpt-oss-20B (high): 0.7%

We evaluate 452 tasks from the public APEX-Agents dataset spanning investment banking, management consulting, and corporate law

Cost (USD) to run the evaluation: GPT-5.4 (xhigh): $1,110, Claude Opus 4.6 (max): $1,055

Corporate Lawyer: Force Majeure Under Executive Order... Management Consultant: 2026 Capital Budget Allocation... Investment Banking Analyst: KVUE DCF Update

Qwen3.5 397B A17B: 15.3%, DeepSeek V3.2: 14.5%, GLM-5: 14.5%, Kimi K2.5: 11.5%, MiniMax-M2.7: 10.6%

APEX-Agents requires agents to navigate realistic work environments with files and tools.

Gemini 3 Flash achieves the highest score of 24.0%

GPT-5.4 (xhigh) scores the highest on APEX-Agents-AA Pass@1 with a score of 33.3%, followed by Claude Opus 4.6 (Adaptive Reasoning, Max Effort) with a score of 33.0%, and Gemini 3.1 Pro Preview with a score of 32.0%

an agent does not care about the structure, unless you specifically ask it to. But even in this case you have to review the changes.

LLMs are pretty good at picking up the style in your repo. So keeping it clean and organized already helps.

When you give a task to your agent, make sure you also explain how the code should be organized. Not only value, but also structure.

Robert Martin in Clean Architecture talks about code as having two properties: value (it works, it's fast, etc.) and structure (how code is organised).

poorly organized code means agents need to read, "understand", and make changes to more files than necessary - polluting their context and costing you tokens.

Context is basically how many things a machine can keep in its operational memory - it's not so different from the very human cognitive load.

their productivity is affected by the state of the codebase.

The human's job is to curate sources, direct the analysis, ask good questions, and think about what it all means. The LLM's job is everything else.

The idea is related in spirit to Vannevar Bush's Memex (1945) — a personal, curated knowledge store with associative trails between documents. The part he couldn't solve was who does the maintenance. The LLM handles that.

Think of fan wikis like Tolkien Gateway — thousands of interlinked pages covering characters, places, events, languages, built by a community of volunteers over years. You could build something like that personally as you read, with the LLM doing all the cross-referencing and maintenance.

Humans abandon wikis because the maintenance burden grows faster than the value. LLMs don't get bored, don't forget to update a cross-reference, and can touch 15 files in one pass. The wiki stays maintained because the cost of maintenance is near zero.

good answers can be filed back into the wiki as new pages. A comparison you asked for, an analysis, a connection you discovered — these are valuable and shouldn't disappear into chat history.

Obsidian is the IDE; the LLM is the programmer; the wiki is the codebase.

the wiki is a persistent, compounding artifact. The cross-references are already there. The contradictions have already been flagged. The synthesis already reflects everything you've read.

Instead of just retrieving from raw documents at query time, the LLM incrementally builds and maintains a persistent wiki — a structured, interlinked collection of markdown files that sits between you and the raw sources.

the maintenance burden grows faster than the value.

what makes the LLM a disciplined wiki maintainer rather than a generic chatbot.

good answers can be filed back into the wiki as new pages.

Obsidian is the IDE; the LLM is the programmer; the wiki is the codebase.

the wiki is a persistent, compounding artifact.

the LLM is rediscovering knowledge from scratch on every question.

design the environment well, you let the agent run, and you own what it produces.

Transparency makes speed feel safe.

Agents should work through the same patterns and actions that humans use.

a stream of text that’s hard to hold onto, hard to compare, and hard to connect

it almost always traces back to the interface rather than the language model

Non-deterministic software breaks the contract.

since reasoning models and agentic AI can rack up quite a bill

Fields that are not exposed now will become exposed in the future

We need, like, a Manhattan Project to collect this

how much demand for something changes when its price changes.

Exposure alone is a completely meaningless tool for predicting displacement

Building a datacenter is supposed to be a “safe” investment in normal times, so banks give private credit and mortgages to finance them.

would shareholders vote to spend 22% of an established company’s market cap to rescue a money-burning AI lab that has lost most of its differentiators?

Raising prices will for sure decrease demand and that risks killing the growth story. And even if revenue keeps growing, it doesn’t matter if there are no margins

They can simply deploy month by month until their competitors struggle to raise and get forced to capitulate. At that point they can just ramp down the spending

they don’t have to spend it to win. It’s a defensive move for them, if they commit $50B, OpenAI and Anthropic need to go raise $100B each to stay competitive

AI is here to stay. If used right, chances are it will make us all more productive. That, on the other hand, does not mean it will be a good investment.

All of this happened in the background. This was just one of the parallel flows in a day. The productivity ceiling? Still unmaxxed.

A fourth built the presentation using a JavaScript library. A fifth critiqued the overall flow & content.

The secret is parallelization. Structure a plan at the start of the day that allows multiple agents to work simultaneously.

The question : how much electricity can we turn into useful work?

That’s up 20x in six weeks. This idea, called tokenmaxxing, is the deliberate practice of maximizing token consumption.

You can’t step outside the forest to warn people about the forest. There is no outside.

AI companies needed human openness to build their models, but will also kill the openness because the relationship is one-sided.

The sheer act of thinking outside the box makes the box bigger.

The platform will know your idea _is pregnant_ far before you will.

The platform doesn’t need to bother with individual prompts - it just needs to see where the questions cluster.

But in the cognitive dark forest, the most dangerous actor is not your peer. It’s the forest itself.

Ideas are cheap - execution is hard -and- the world ahead is ripe with opportunity.

A learning system can continuously incorporate real-world data in a way that numerical solvers fundamentally cannot, capturing and compounding the knowledge that is currently trapped out there in the real world.

Worse, they learn nothing from past work. Institutional knowledge lives in textbooks and the minds of a few experts. None of it is captured in the tools themselves.

They meet their target S-parameter specifications despite having very alien-looking geometries.

Learning fields turns S-parameter extrapolation into something closer to an in-distribution task.

Training on fields themselves forces the model to learn the physics that produces S-parameters, rather than learning to approximate the mapping directly.

A wire becomes a transmission line. A bend becomes a reflector. Two parallel traces become coupled antennas. The geometry is the circuit.

identify if the app had been left in a broken state

but would fail recognize that the feature didn't work end-to-end

commit its progress to git with descriptive commit messages

inappropriately change or overwrite JSON files compared to Markdown files

see that progress had been made, and declare the job done

each new engineer arrives with no memory of what happened on the previous shift

harness combinations doesn't shrink as models improve. Instead, it moves

a harness encodes an assumption about what the model can't do on its own