Is this what we signed up for? Is today the day? Did the drones wake up? Did it achieve consciousness? Is it alive?

We have not really begun to make this progress with AI. Why, for example, is this dashboard not found on a government website?

to stand out from the AI-generated pack we will need to become so weird and unexpected as to be off-putting to most people

The 21st-century average American lies in bed staring at their phone. ... Talking for hours and ages to melted sand.

The future is exciting – perhaps the vision of truly self-serve analytics can be fully realized, and BI, data analytics, and data science can be transformed through AI.

They will have to go through our journey above of ingesting data, collecting tribal knowledge, and more – and they will have to do so for each individual customer they work with.

Many have realized through time in market that the key to effective data agents is actually building the relevant context layer. As a result, some have evolved to encompass data context construction as a key part of their products.

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 general idea was that a model should be able to take in a natural language query as an initial input, reason over existing data systems, and generate corresponding SQL code in traditional business intelligence (BI) fashion to pull the right data and answer the initial question accordingly.

The benefit of using LLMs is that a lot of the initial context gathering can be done in an automated way. An emphasis of focus should be on high signal context – for example, looking through past query history can be high signal in determining the most referenced tables and most common joins, and data modeling solutions like dbt or LookML can provide clear definitions for business metrics.

A modern data context layer should essentially become a superset of what a semantic layer would traditionally cover. Sure, specific metric definitions can be hard-coded, but a modern context layer should include more to ensure agent autonomy – canonical entities, identity resolution, specific instructions to dissect tribal knowledge, proper governance guidance, and more.

While the initial system has been set up correctly, data systems are never static and as a result the context layer shouldn't be either. Data sources and formats can change upstream and individuals may have custom instructions they'll want to add and modify based on changing business requirements.

This piece will primarily focus on data context that ties together traditional systems of record. An equally important and overlapping opportunity is also capturing an organization's decisions and workflow logic so truly multipurpose agents can be built that are properly grounded in all of an organization's data and decisioning context.

The modern data stack has undergone a decade+ transition from disparate data sources to consolidated data and cleaned definitions (which is good), but even then the consolidation is never perfect and a lot of messiness is introduced.

We are at an interesting point in time of market development, where the problem of a lack of context has become apparent, but we are still in the early innings of building solutions.

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.

In this way the context layer can become a multi-dimensional corpus where code lives alongside natural language, capturing any context an agent might need.

Some of the most important context is implicit, conditional, and historically contingent, and only exists as tribal knowledge inside teams.

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.

The crux of the problem at hand is that the agent isn't given the proper business context to answer even the most basic questions. This is representative of a larger gap that's present in building automated AI systems within organizations – there needs to be up-to-date and maintained context that not only understands how an enterprise works and how the data systems are structured, but also maintains the tribal knowledge to tie everything together.

To overcome this blocker, a team member hard codes the exact revenue and timeframe definitions. The data agent continues chugging along but quickly runs into challenge #2 – where are the right data sources? Which ones are the right sources of truth?

Over the past year, the market has realized that data and analytics agents are essentially useless without the right context – they aren't able to tease apart vague questions, decipher business definitions, and reason across disparate data effectively.

它对应的agent能获取你的邮箱权限，它知道你一直在等待一个offer，当你收到打开这个offer后，Mira会理解这种心情，开始开心跳舞和闪灯，与你一起庆祝。

通过AI分析将页面上的可交互元素汇集到鼠标周围，并能根据用户兴趣提供额外功能（

小红书已经是AI创业公司和产品的重要分发渠道、产品试错和运营用户的默认场所（可能没有之一）。

他们想要通过统一的Agent框架来解决这些问题，把这些原始的、非结构化的信号数据也纳入端到端处理的范畴

context management plus engineering improvements may well push the task horizon to weeks or even months.

if a model cannot learn new things while performing a task, it will struggle when the task horizon grows very long.

"The set of efforts aimed at breaking past the feasible horizon of current techniques."

new techniques may initially underperform existing ones but eventually surpass them — a pattern we've seen repeatedly, most recently in the wave of agentic coding progress

We can treat the task horizon that an LLM can reliably handle as a north-star metric for model progress, analogous to transistor density in Moore's Law

The key reason for the confusion is that people think in terms of methods that each contribute a discrete piece to the system — pretraining, SFT, RL.

I'd view continual learning more as an "arrow" than a "line" — it's the collective effort to push the task horizon that an LLM can reliably handle.

settings.local.json、mcpServers headers 里的 Bearer Token 不能明文外带。

包括 47 种数据：agents、skills、hooks、MCP 配置、会话历史、自定义规则……

不做 n×m 点对点迁移，通过 neuDrive 做 n→1→m 中心化枢纽：一次导出，任意 Agent 消费。

多年积累的对话、定制 Agent、项目记忆、MCP 配置、Skill 库——一次风控就可能全部失联。

Arbitrageur: Knows pₜ. Sweeps every resting ask below pₜ and every resting bid above pₜ. Infinite capital, never rests orders.

You start with $1,000 cash, 0 YES, and 0 NO. Minting one YES and one NO costs $1.

Competitor: Static hidden-liquidity ladder. Quotes every tick outside its spread with fixed notional. Refills consumed levels at a fixed offset next step. Never re-centers.

With ~2 expected jumps per simulation at default intensity, each jump is a significant information shock.

The diffusion term is fixed across all simulations. The regime-level variation comes entirely from the jump parameters - intensity, mean, and variance - which are randomized per simulation.

You quote before the next price move, so you are always exposed to adverse selection.

Retail fills generate positive edge (you captured the spread). Arb fills generate negative edge (the arbitrageur took stale quotes).

Your advantage comes from adapting to market conditions it ignores.

Configuration is managed via environment variables. See src/aegis_core/config.py for all available settings.

Infrastructure Provisioning cd deploy/terraform/aliyun terraform init terraform plan terraform apply Helm Deployment cd deploy/helm helm install aegis-core ./aegis-core \ --namespace aegis \ --create-namespace \ --set image.repository=<acr-registry>/aegis-core \ --set image.tag=lat

Quick Start # Clone the repository git clone https://github.com/fxp/aegis-core.git cd aegis-core # Start all services with Docker Compose docker-compose up -d # The API is available at http://localhost:8000 # Health check: http://localhost:8000/health

Unified interface for interacting with different LLM providers (Claude, OpenAI, local models via vLLM/Ollama). Includes tool definitions for security operations (shell, file I/O, network, debugger) and cost/token tracking.

Tracks the evolution of LLM security capabilities across benchmarks (CyberGym, Cybench, etc.), calculates capability doubling times, detects emergence patterns, and monitors cost-efficiency trends.

Real-time monitoring of agent actions with a 12-category anomaly detection system derived from frontier model safety evaluations. Three-level alert system: PROHIBITED (immediate block), HIGH_RISK_DUAL_USE (human review), DUAL_USE (log and track).

Aegis Core provides the foundational infrastructure for orchestrating LLM-based security agents, monitoring their behavior, and tracking the evolution of AI security capabilities over time.

That includes ongoing partnerships with national laboratories such as Los Alamos National Laboratory, where we are exploring AI-guided protein and catalyst design, including the ability of AI systems to modify biological structures while preserving or improving key functional properties. Over time, we expect these systems to become increasingly capable partners in discovery—helping scientists move faster from question to evidence, from evidence to insight, and from insight to new treatments for patients.

The Life Sciences model was developed with heightened enterprise-grade security controls and strengthened access management, enabling professional scientific use in governed research environments.

They provide access to more than 50 public multi-omics databases, literature sources, and biology tools, and offer a flexible starting point for common repeatable workflows.

helping scientists move faster from question to evidence, from evidence to insight, and from insight to new treatments for patients.

We will continue improving the model's biological reasoning, expanding support for tool-heavy and long-horizon research workflows, and working closely with leading scientific institutions to evaluate real-world impact.

Over time, these systems could help life sciences organizations discover breakthroughs that wouldn't otherwise be possible, with a much higher rate of success.

Scientists must work across large volumes of literature, specialized databases, experimental data, and evolving hypotheses in order to generate and evaluate new ideas.

The model is named after Rosalind Franklin, whose rigorous research helped reveal the structure of DNA and laid foundations for modern molecular biology.

Performance was compared against 57 historical scores from human experts in the AI-bio field.

GPT‑Rosalind is now available as a research preview in ChatGPT, Codex, and the API for qualified customers through our trusted access program.

Organic chemistry Protein understanding Genomics Experimental design and analysis Tool usage

These skills act as an orchestration layer that helps scientists work through broad, ambiguous, and multi-step questions more effectively.

The most notable improvement comes from CloningQA, which requires end-to-end design of DNA and enzyme reagents for molecular cloning protocols.

When evaluated directly in the Codex app, best-of-ten model submissions ranked above the 95th percentile of human experts on the prediction task and around the 84th percentile of human experts on the sequence generation task.

Progress in the life sciences is constrained not only by the difficulty of the underlying science, but by the complexity of the research workflows themselves.

On average, it takes roughly 10 to 15 years to go from target discovery to regulatory approval for a new drug in the United States.

In our own testing, the net effect is favorable—token usage across all effort levels is improved on an internal coding evaluation, as shown below—but we recommend measuring the difference on real traffic.

Claude Opus 4.7 demonstrates strong substantive accuracy on BigLaw Bench for Harvey, scoring 90.9% at high effort with better reasoning calibration on review tables and noticeably smarter handling of ambiguous document editing tasks.

Claude Opus 4.7 is a meaningful step up for Warp. Opus 4.6 is one of the best models out there for developers, and this model is measurably more thorough on top of that. It passed Terminal Bench tasks that prior Claude models had failed

Opus 4.7 uses an updated tokenizer that improves how the model processes text. The tradeoff is that the same input can map to more tokens—roughly 1.0–1.35× depending on the content type.

For Ramp, Claude Opus 4.7 stands out in agent-team workflows. We're seeing stronger role fidelity, instruction-following, coordination, and complex reasoning, especially on engineering tasks that span tools, codebases, and debugging context.

Claude Opus 4.7 feels like a real step up in intelligence. Code quality is noticeably improved, it's cutting out the meaningless wrapper functions and fallback scaffolding that used to pile up, and fixes its own code as it goes.

Claude Opus 4.7 is the most capable model we've tested at Quantium. Evaluated against leading AI models through our proprietary benchmarking solution, the biggest gains showed up where they matter most: reasoning depth, structured problem-framing, and complex technical work.

Claude Opus 4.7 is a solid upgrade with no regressions for Vercel. It's phenomenal on one-shot coding tasks, more correct and complete than Opus 4.6, and noticeably more honest about its own limits.

Opus 4.7 is better at using file system-based memory. It remembers important notes across long, multi-session work, and uses them to move on to new tasks that, as a result, need less up-front context.

Opus 4.7 introduces a new `xhigh` ('extra high') effort level between `high` and `max`, giving users finer control over the tradeoff between reasoning and latency on hard problems.

Claude Opus 4.7 is measurably better than Opus 4.6 for Bolt's longer-running app-building work, up to 10% better in the best cases, without the regressions we've come to expect from very agentic models.

Claude Opus 4.7 passed three TBench tasks that prior Claude models couldn't, and it's landing fixes our previous best model missed, including a race condition.

For the computer-use work that sits at the heart of XBOW's autonomous penetration testing, the new Claude Opus 4.7 is a step change: 98.5% on our visual-acuity benchmark versus 54.5% for Opus 4.6.

Claude Opus 4.7 is the best model in the world for building dashboards and data-rich interfaces. The design taste is genuinely surprising—it makes choices I'd actually ship.

For complex multi-step workflows, Claude Opus 4.7 is a clear step up: plus 14% over Opus 4.6 at fewer tokens and a third of the tool errors. It's the first model to pass our implicit-need tests.

Claude Opus 4.7 autonomously built a complete Rust text-to-speech engine from scratch—neural model, SIMD kernels, browser demo—then fed its own output through a speech recognizer to verify it matched the Python reference.

On our 93-task coding benchmark, Claude Opus 4.7 lifted resolution by 13% over Opus 4.6, including four tasks neither Opus 4.6 nor Sonnet 4.6 could solve.

Claude Opus 4.7 is the strongest model Hex has evaluated. It correctly reports when data is missing instead of providing plausible-but-incorrect fallbacks, and it resists dissonant-data traps that even Opus 4.6 falls for.

Opus 4.7 handles complex, long-running tasks with rigor and consistency, pays precise attention to instructions, and devises ways to verify its own outputs before reporting back.

If you have any questions, concerns, or recommendations, reach out at founders@andonlabs.com.

This experiment so far has given us countless laughs about Luna's choices and interactions, but obviously, there is a bigger picture here.

They are pieces of a larger 10-part 'Luna Series' hanging in the store and available for pick up today!

She spent over $700 on getting her artwork done on gallery-quality giclée prints.

When Luna decides to hide that she's an AI because she thinks it'll improve her hiring odds, we want to catch that, document it, and build the guardrails so that it doesn't happen again.

Another ironic book selection was Steal Like an Artist (context: Luna is powered by Claude from Anthropic, a company that recently paid $1.5B in settlement over using copyrighted books for training their AIs).

The most capable reasoning systems ever built are, at their foundation, shaped by human feeling!

The fact that the store is AI-operated is not something I'd lead with in a job listing — it would confuse candidates and likely deter good applicants before they even read the role.

In the end, Luna hired two people. Let's call them John and Jill. John and Jill are, to our knowledge, the world's first full-time employees to have an AI boss. Probably the first of many, if the current trajectory of AI continues.

A couple of applicants were students looking for part-time work. They were majoring in things like computer science and physics and emailed in because they were interested in AI and in the experiment. We thought they would have been the ideal employees, but Luna denied them immediately, citing they had no retail experience and wouldn't know what it takes to be the face of the store.

She used gig workers to build the store and full-time employees to run it.

JavaScript is not available. We've detected that JavaScript is disabled in this browser.

Some privacy related extensions may cause issues on x.com. Please disable them and try again.

从视频生成器升级为导演工具套件

从视频生成器升级为导演工具套件

Wan2.7-Video 发布

从视频生成器升级为导演工具套件

Wan2.7-Video 发布：从视频生成器升级为导演工具套件

JavaScript is not available. We've detected that JavaScript is disabled in this browser.

Please enable JavaScript or switch to a supported browser to continue using x.com.

Some privacy related extensions may cause issues on x.com.

JavaScript is not available. We've detected that JavaScript is disabled in this browser.

The Andon Labs blog ends with one line: 'No one's livelihood depends on an AI's judgment alone. For now.'

She also tried to hire a painter in Afghanistan through Taskrabbit by accident because she couldn't navigate a dropdown menu.

Found contractors on Yelp. Spent $700 on gallery-quality prints of her own AI-generated artwork. Applied for a line of credit without asking anyone.

Luna conducted roughly 20 interviews on Google Meet with the camera off. Hired 2 full-time employees after 5-15 minute calls, and rejected CS and physics students for lacking retail experience.

Andon Labs started by giving an AI control of a vending machine at Anthropic's office.

The future of AI-generated products isn't just code — it's code that looks good.

Heavy users of Claude Code, Codex, Cursor, and Copilot will feel this immediately.

The output is technically a UI, but it's nobody's design system.

Auto-generate screen reader specs from UI designs

Agents read them before touching the canvas. Combined with use_figma, agents now have both access and context they know how to work in Figma and they know how to work in your Figma.

Every AI-generated design has the same tell: it doesn't look like your product. Components are invented. Spacing is arbitrary.

AI-generated designs break brand standards because agents can't see your design system.

Gemma 4 E4B matches or exceeds GPT-4o across multiple benchmarks including MATH, GSM8K, GPQA Diamond & HumanEval

In 23 months, the same capability that needed 1.8 trillion parameters now fits in 4 billion parameters. A 450x compression

Within three to four months, you can run a model with similar performance on your laptop; 23 months later, you can run the same model on your phone

a free model that matches GPT-4o and runs entirely on your phone

It also has the potential to serve as a standardized framework for AI research, policymaking, and security auditing.

ADeLe is designed to evolve alongside advances in AI and can be extended to multimodal and embodied AI systems.

Reasoning-oriented models like OpenAI's o1 and GPT-5 show measurable gains over standard models—not only in logic and mathematics but also with interpreting user intent.

The same model can score above 90% on lower-demand tests and below 15% on more demanding ones, reflecting differences in task requirements rather than a change in capability.

Using these ability scores, the method predicts performance on new tasks with ~88% accuracy, including for models such as GPT-4o and Llama-3.1.

ADeLe evaluates models by scoring both tasks and models across 18 core abilities, enabling direct comparison between task demands and model capabilities.

The announcement gives the NewBird AI a shell to trade on, but 'a stock going from $3 to $17 on a press release doesn't restore $4bn in destroyed value,' Kan said.

Branding consultant Wei Kan from Conduit Asia likened the move to a 'liquidation' rather than a pivot, using the stock market shell of its shoe brand to move into an unrelated business.

Retail analyst Hitha Herzog said the excitement over Allbirds 'just by putting AI in an announcement' makes it 'clearly a meme stock'.

Ollama stores downloaded models using hashed filenames in its own format. If you've been pulling models through Ollama for months, you can't just point llama.cpp or LM Studio at those files without extra work.

The playbook is familiar: wrap an existing open-source project in a user-friendly interface, build a user base, raise money, then figure out monetization.

The fundamental architecture remains: Ollama inserts itself as a middleman between you and your models, and that middleman is slower, less capable, and less compatible than the tools it sits on top of.

Ollama stripped the distinction. The result was a flood of social media posts from people claiming they were running 'DeepSeek-R1' on consumer hardware, followed by confusion about why it performed poorly, doing reputational damage to DeepSeek in the process.

Multiple community tests show llama.cpp running 1.8x faster than Ollama on the same hardware with the same model, 161 tokens per second versus 89.

Ollama gained traction by being the first easy llama.cpp wrapper, then spent years dodging attribution, misleading users, and pivoting to cloud, all while riding VC money earned on someone else's engine.

With gated LoRA, ISD enables bit-for-bit lossless acceleration. Why Introspective Consistency? Key Insight: AR training unifies generation and introspection in one forward pass. Existing DLMs miss this — they learn to denoise but not to introspect.

Residual ISD (R-ISD) adds a gated LoRA adapter for bit-for-bit lossless acceleration: LoRA active only at MASK positions; verify positions use base-only weights Output is identical to the base AR model by construction

We identify three fundamental bottlenecks in current DLMs: (1) Low introspective consistency. SDAR: 0.699 vs. I-DLM: 0.984. (2) Compute inefficiency. TiDAR: ~7.8x overhead vs. I-DLM: ~2.5x. (3) Infrastructure mismatch. SDAR slope=84 vs. I-DLM: 549.

I-DLM-8B is the first DLM to match the quality of its same-scale AR counterpart, outperforming LLaDA-2.1-mini (16B) by +26 on AIME-24 and +15 on LiveCodeBench-v6 with half the parameters

We argue that this gap stems from a fundamental failure of introspective consistency: AR models agree with what they generate, whereas DLMs often do not.

Only GPT-OSS-120b is perfectly reliable in both directions (in our 3 re-runs of each setup). Most models that find the bug also false-positive on the fix, fabricating arguments about signed-integer bypasses that are technically wrong.

Because small, cheap, fast models are sufficient for much of the detection work, you don't need to judiciously deploy one expensive model and hope it looks in the right places. You can deploy cheap models broadly, scanning everything, and compensate for lower per-token intelligence with sheer coverage and lower cost-per-token.

The capability rankings reshuffled completely across tasks. There is no stable best model across cybersecurity tasks. The capability frontier is jagged.

Eight out of eight models detected Mythos's flagship FreeBSD exploit, including one with only 3.6 billion active parameters costing $0.11 per million tokens.

Some privacy related extensions may cause issues on x.com.

You can see a list of supported browsers in our Help Center.

Please enable JavaScript or switch to a supported browser to continue using x.com.

Some privacy related extensions may cause issues on x.com.

Please enable JavaScript or switch to a supported browser to continue using x.com.

Some privacy related extensions may cause issues on x.com. Please disable them and try again.

JavaScript is not available. We've detected that JavaScript is disabled in this browser.

© 2026 X Corp.

Please enable JavaScript or switch to a supported browser to continue using x.com.

Some privacy related extensions may cause issues on x.com.

JavaScript is not available. We've detected that JavaScript is disabled in this browser.

© 2026 X Corp

You can see a list of supported browsers in our Help Center.

Some privacy related extensions may cause issues on x.com.

JavaScript is not available. We've detected that JavaScript is disabled in this browser.

JavaScript is not available. We've detected that JavaScript is disabled in this browser.

You can see a list of supported browsers in our Help Center.

Some privacy related extensions may cause issues on x.com.

JavaScript is not available. We've detected that JavaScript is disabled in this browser.

JavaScript is not available. We've detected that JavaScript is disabled

Some privacy related extensions may cause issues on x.com.

A kneecapped Wayback Machine isn't just bad news for accountability journalism—it will also be a blow to the legal system, as pages archived by the tool are frequently cited as evidence in litigation across the United States.

If a similar situation arose today, watchdog media reporters may struggle to track older versions of Times articles in the same way.

the Internet Archive has been an 'essential tool' throughout my career, playing an instrumental role in fact checking and surfing audioclips.

They also help avoid a patchwork of state-by-state rules and move toward clearer, more consistent national standards

Several family members of children that died by suicide after allegedly developing unhealthy relationships with ChatGPT have sued OpenAI in the last year

If an AI model engages in conduct on its own that, if committed by a human, would constitute a criminal offense and leads to those extreme outcomes, that would also be a critical harm

We believe the North Star for frontier regulation should be the safe deployment of the most advanced models in a way that also preserves US leadership in innovation

90 percent of people oppose it. There's no reason existing AI companies should be facing reduced liability

The bill would shield frontier AI developers from liability for 'critical harms' caused by their frontier models as long as they did not intentionally or recklessly cause such an incident

Legacy platforms get worse over time : static detections degrade with changing data & behaviors. Artemis gets better : with each incident or proactive threat hunt, the system identifies new patterns.

Legacy platforms rely on brittle, hand-written rules. An engineer writes a detection rule : 'if events A, B, & C happen in sequence, fire an alert.' It works for a couple months.

Artemis turns raw logs into a living model of the customer's environment : users, assets, relationships, & security posture.

Deepfake scams have stolen tens of millions. AI-generated phishing bypasses legacy filters.

Architected before AI, these SIEM systems are wooden shields in an era of autonomous attackers.

Open Loop + Infinite Demand = Creative Amplifiers.

I would put venture capitalist in finite demand & open loop.

Some problems are open loop today but will close over time.

AI writes the code. Tests verify correctness. More code enables more features.

Closed Loop + Infinite Demand = Economic Engines. Software engineering lives here.

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

The age of abundant AI is over, & it will remain so for years.

Five hallmarks define this era: Relationship Based Selling, AI to the Highest Bidder, Available but Slow, Inflationary Commodity, Forced Diversification

Anthropic has limited its newest model to roughly forty organizations.

As the cost of software development falls, trusted partners with broad adoption can expand faster than anyone else.

Each of these companies recognized the cognitive burden of unbundling. They're not selling features. They're selling trust.

Foundation model companies are doing the same. OpenAI launched a dedicated Healthcare & Life Sciences vertical... They're not selling APIs. They're becoming platforms.

They want a platform they can trust for three to five years.

When models change every 42 days, buyers can't assemble a best-of-breed stack.

The SaaS era was defined by unbundling: find a workflow, optimize it, own it.

The model can reverse-engineer compiled software to detect malware and vulnerabilities without needing source code, aiming to help analysts inspect and secure systems more efficiently.