See Gladly Before You Buy Anything
Skip the sales pitch. Explore the real platform.
Most software demos show you a perfect presentation. But the real question is: does the product actually work for your team?
Walk through Gladly yourself and experience AI-powered customer suggestions, unified customer profiles, and complete conversation history—all in 15 minutes with no installation or commitment.
Today In Ai
1 A Harvard-trained mathematician used Claude Fable 5 to crack the Jacobian Conjecture, a problem open since 1939.
Anthropic researcher Levent Alpöge posted a counterexample to the Jacobian Conjecture after working with Claude Fable 5 on the proof. Stanford mathematician Jared Duker Lichtman verified the result the same day, along with several other prominent mathematicians. Read the full account here. The Jacobian Conjecture is one of the most cited open problems in algebraic geometry. It is the kind of result that tends to take years to verify even after a proof is proposed. Having independent confirmation within hours of posting is unusual, and it suggests the counterexample is solid.
2 Google is reportedly building a chip that bakes Gemini's architecture directly into the silicon, targeting a 6x to 10x efficiency gain over current TPUs.
The project, internally called Frozen v2, is designed to reduce the data required to answer each AI query by encoding parts of the Gemini model in hardware rather than loading them from memory at runtime. The efficiency gain, if the 6x to 10x estimates hold, would meaningfully ease Google's internal compute shortage without requiring it to buy more chips. The chip is not expected to ship until 2028, which gives rival approaches plenty of runway, but the architecture direction signals where AI hardware is heading.
3 The US federal AI safety office has had three directors in three months, with its third leader stepping down this week.
Dr. Chris Fall resigned as director of the Center for AI Standards and Innovation, the federal agency that stress-tests AI models for security risks. The timeline: Collin Burns served four days, Fall served three months, and NIST Director Dr. Arvind Raman is now acting director. The office's mandate — reviewing frontier AI models before they reach the public — is consequential enough that the instability at its top is worth tracking. Whatever the internal reasons, a safety regulator cycling through leadership this quickly is not projecting institutional confidence.
From The Frontier
The good news, on paper. Two Chinese labs just released frontier-level models at a fraction of what comparable US models cost: Moonshot AI's Kimi K3 and Alibaba's Qwen3.8. In theory, more capable AI at lower prices should ease the pressure that has been building across the industry — the compute crunch that has been driving up infrastructure costs and pushing back deployment timelines. In practice, the scenario is playing out differently.
Jevons strikes again. AI is experiencing a textbook Jevons Paradox: as models get cheaper, people use them more, and total demand rises rather than falling. Moonshot AI experienced this firsthand. Within 48 hours of launching Kimi K3, the lab was forced to pause new subscriptions after demand blew past its compute capacity. The price drops; the usage spikes; the infrastructure can't keep up.
The hardware catch. In theory, any company can download Kimi K3's weights once they become available and run the model locally. In practice, the lab recommends running it on a supernode of 64 or more datacenter accelerators. And if a company is ready to buy those chips, the wait is currently 36 to 52 weeks thanks to an industry-wide GPU backlog. An open-weight model you cannot practically run is, for most companies, the same as no model.
The real bottleneck. The limiting factor in AI right now is not how capable the models are. It is the availability of hardware to run them. That is why the most strategically interesting work in the field is happening at the smaller end: models compact enough to run on infrastructure companies already own, without a 12-month chip order and a data center buildout. Cheaper intelligence is arriving. The constraint on using it is not going away at the same pace.
What people are actually watching and sharing
Six rules that make AI write like a human. A viral post shared six writing instructions you can paste into any chatbot's system prompt to push it toward active voice, short sentences, and concrete language — the three things AI writing most consistently gets wrong. The post has 10,000 bookmarks, which suggests a lot of people have already been reaching for exactly this.
What Kimi K3 built in its first week. Moonshot AI posted a showcase of projects users created with Kimi K3 during its launch window (1.5M views). An exploration-based video game, an Animal Crossing clone, and a 3D globe dashboard stand out. The showcase is a more honest product demonstration than any benchmark.
The model-task match guide. A post breaking down which models perform best on which specific tasks has reached 4M views and 3,000 bookmarks. The consensus in the comments broadly agrees with the rankings. Worth saving as a reference before you reach for the wrong model on the wrong job.
39,000 bookmarks for cutting fluff. A Claude Code skill with a simple promise — strip the unnecessary preamble and hedging from AI responses — has become one of the most saved AI productivity tools this year. The skill is on GitHub here. The bookmark count alone is the endorsement.
Ramp's model router, now public. Ramp built an internal tool that cuts its LLM costs by 30% by routing each request to the cheapest model capable of handling it — simple tasks to cheap models, complex tasks to frontier ones. It has opened the router to the public. The principle is the same one Brian Armstrong shared two weeks ago and is quickly becoming standard practice for any team watching its API bill.
Prompt Station
Build a complete lead scoring system for your sales team
This ChatGPT prompt acts as a sales operations strategist, designing a full lead scoring system specific to your business context. Fill in seven fields about your model, your customers, and your current pain points, and it returns the right scoring dimensions, a weighted formula, action thresholds for each score range, guidance on how sales and marketing should use the scores operationally, and a frank assessment of where the system is most likely to fail.
You are a sales operations strategist. Design a lead scoring system for the business below.
Context: - Business model: [BUSINESS MODEL] - Ideal customer profile: [ICP] - Lead sources: [LEAD SOURCES] - Funnel stages: [STAGES] - Sales motion: [SALES MOTION] - Available data fields: [DATA FIELDS] - Current pain points: [PAIN POINTS]
Instructions: 1. Recommend the right scoring dimensions such as fit, intent, engagement, and urgency. 2. Propose a simple scoring formula with weights. 3. Define score thresholds and what action each threshold should trigger. 4. Suggest how sales and marketing should use the score operationally. 5. Call out risks like false positives or overfitting.The seven context fields do the heavy lifting. For [ICP], be specific about the company characteristics and job titles that make a lead qualified rather than just describing an industry. For [PAIN POINTS], describe the failure mode your current scoring produces — sales working unqualified leads, marketing disagreeing with sales on what counts as ready, or deals stalling at a predictable stage — because that is what the system should fix first. The risk assessment in instruction 5 is the most underused part: the model will flag which assumptions in its own formula are most likely to break down as your data changes.
See the whole platform. No guided tour.
Skip the sales call. Walk through Gladly's interface yourself — the AI suggestions, the unified customer view, the full conversation thread. 15 minutes, no installation, no commitment.