We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
1. Funding & Deals
The Horowitz Andreessen Academy disclosed $42M in total funding led by a16z, without naming a financing stage. It is a selective San Francisco school for young builders; its first offering is a tuition-free, one-year fellowship scheduled to begin in Fall 2027. Ten founding partners—including Anduril, Anthropic, Google, NVIDIA, and OpenAI—are expected to contribute compute or hardware, curriculum input, and co-op opportunities; the launch also lists 50+ hiring partners. The launch interview describes the incoming CEO as a decade-plus entrepreneur who co-founded Udemy and Maven. This is best read as a bet on AI-era talent formation and employer access, rather than a conventional AI-product financing.
2. Emerging Teams
Andy has come out of stealth with a network-first wedge into fragmented corporate entertainment. Founder Loit Sarma co-founded an ADP-backed HR venture and spent 2.5 years at ADP overseeing money movement and payments. He says Andy has contracts with 60 large companies, more than 1,600 hospitality groups, and 93,000 venues; case studies claim 12–15% savings, while the roughly $400M of platform volume projected by year-end remains a forecast. Its AI is positioned as a concierge and response layer: the company says humans remain involved and agents are not yet negotiating on customers’ behalf. Diligence should focus on money movement, tax jurisdictions, venue onboarding, and activating additional cities.
3. AI & Tech Breakthroughs
Xiaomi’s MiMo-V2.6-Pro and Flash extend the open-weight contest into natively omnimodal models. Artificial Analysis, as relayed by AINews, ranked Pro first among open-weight models on its Intelligence Index (46), at $0.13 per index task; the report gives 1.02T total and 42B active parameters. The reported asynchronous RL stack spans coding, agents, visual work, and cyber tasks. Xiaomi says it will open-source environment code and training recipes, but not the complete 7,000+ task dataset. AINews cites $2.6M as a conditional estimate for a 130-hour, 75B-token RL run—not an all-in training-cost disclosure. The diligence questions are whether the open toolchain can produce repeatable gains and what costs sit outside that RL run.
World Labs is preparing Atlas for product release as a controllable 3D world model. Fei-Fei Li describes combining camera control and pixel consistency with image generation and reconstruction of intricate 3D structure. She says Atlas outperforms specialized state-of-the-art models, but offers no quantitative comparison in the interview. The application she emphasizes is creating realistic simulations to train and evaluate robots in data-scarce settings such as pharmaceutical manufacturing and laboratory science; she also argues that world-model benchmarking needs independent academic and public-sector participation.
Agent reliability is improving through post-training, but automated evaluation still misses product judgment. Perplexity says its Computer agent learned from real user sessions by imitating good trajectories and correcting avoidable tool-call errors; a live A/B test reported 21.2% fewer tool-call failures than an earlier checkpoint. In a separate analysis of 100 production traces from an apartment-leasing assistant, automated evaluators caught obvious contradictions but missed context-dependent product failures and sometimes flagged good responses. The authors recommend human review and active sampling, making the feedback and adjudication loop—not just the model—the diligence target.
4. Market Signals
A curated VC panel favors AI infrastructure over new foundation-model entrants. Newcomer’s first sentiment report draws on interviews with 25 investors whose firms collectively manage more than $100B. Its published excerpt describes respondents as bullish on AI infrastructure but soured on new model companies beyond the leading labs; they point to billion-dollar entry valuations, incremental Transformer variants, and the need for model vendors to keep delivering strong new releases. Treat this as a qualitative signal from a selected panel, not a representative measure of capital allocation.
The price/performance bar is moving down, and token prices alone are a poor underwriting metric. OpenAI says GPT-6 Sol and Luna bring many of Astra’s strengths into faster, more affordable models, with API prices 50% below GPT-5.6 promotional pricing; Sam Altman says they cost half as much per token and less per task. Separately, Epoch AI estimates that cost at a given performance level has fallen about 47% per quarter since 2023. These are company and research-body claims, not verified margin data. For diligence, compare realized cost per successful task and the usage response; Altman himself argues that per-task pricing is the relevant comparison.
5. Worth Your Time
- Watch — Fei-Fei Li on Atlas. The most useful passage connects world models to realistic robotics training and evaluation environments for data-scarce lab and pharmaceutical work, then makes the case for independent benchmarking.
- Read — Lenny’s “Advanced evals.” Practical on why teams should inspect production failures before writing metrics, and where automated trace review still needs human judgment.
Bottom line: The supplied official homepage supports a featured network, corporate-partner claims, and a hands-on San Francisco program, but does not verify the Academy’s own leadership, founding-team background, or financing. On its own, it is insufficient to substantiate an investment signal.
- Named people: The page lists figures including Sam Altman, Jensen Huang, and Fei-Fei Li under “Learn from the people shaping the future,” with their outside-company roles, and links to “Explore the network.” This is evidence of a featured network, not identification of Academy executives.
- Founding team and partners: The site describes “founding partners” as helping co-create curriculum, provide hardware and compute, and share networks and knowledge; it names corporate partners including Anduril, Anthropic, Google, NVIDIA, and OpenAI. These are partner claims, not biographies or backgrounds of the Academy’s own founding team. The supplied homepage does not disclose an Academy funding round, investors, amount, or terms.
- Program model: The homepage presents a San Francisco-based “Build, work, and learn” model, with a Mon/Wed/Fri schedule strip and examples including building a humanoid robot, a fireside chat, and independent study. It also says students live alongside classmates in San Francisco. Hiring partners provide opportunities for students to be considered for internships and full-time roles—not a stated guarantee of placement.
The supplied excerpt does not verify a 100-production-trace comparison: it introduces the question of what trace-analysis agents catch or miss, but provides no sample count, comparative results, or failure categories.
- The article frames production-trace review as error analysis to catalog product failures, and notes that reviewing traces can change the criteria used to grade outputs (“criteria drift”).
- It says Braintrust Loop, Arize Alyx, and LangSmith Engine read production traces and report failures, but gives no evidence here about which failure types they detect or miss.
- The only concrete human-review workflow detail in this excerpt is that Alyx could not write its analysis back to the dataset in the requested format, so the results were copied into a spreadsheet for manual review. This is an observed workaround, not a complete recommended workflow.
Direct answer: This source does not verify the substantive RL-training claims or explain what the reported training cost includes. It identifies the repository as XiaomiMiMo/MiMo-V2.6-Pro-RL and links a 3.05 MB technical-report PDF, but the supplied extract does not include the report’s contents or a cost breakdown.
- The extract supplies no verifiable model parameter count or RL task/training setup.
-
The page has
Transformers,Safetensors, andcustom_codetags, but those tags and the shown report-PDF link do not establish which training code or data were released.
- a16z announced an academy as a separate company rather than part of its investment firm. It is framed as a non-degree, residential San Francisco program for high-school students, with employer connections and a college-comparable experience; its approach emphasizes project work and self-paced learning with AI tools over lecture-heavy instruction.
- The discussion presents Gagen as the academy’s co-founder and CEO choice, citing his Udemy co-founding, more recent Maven experience, and decade-plus entrepreneurial track record. a16z positions its portfolio of 1,600 companies as a resource and says it is working to make graduate opportunities with companies explicit.
- The planned founding-class fellowship is for 50 students; leadership says it sees clear demand among young builders, based on meeting them over the prior six months. The speakers say earlier efforts in the space did not break through and argue that a competitive offering must bundle learning with credential/status/brand, social experience, and employer links—not just narrow job training.
- Fei-Fei Li’s World Labs is working toward releasing Atlas as a product. The company describes the world model as combining generated pixels, camera control, and reconstruction of intricate 3D structure; Li says it outperforms state-of-the-art specialized models on these tasks.
- Li says World Labs is engaging users in creative, design, robotics, healthcare, and pharmaceutical-lab sectors. For robotics, she says world models could create realistic simulated training and evaluation environments, addressing bottlenecks in specialized settings where data are scarce; she also argues that model evaluation and benchmarking require ecosystem-wide participation, including academia, the public sector, industry, and government.
- Alibaba unveiled an AI accelerator aimed at frontier-model training and competition with Nvidia and Huawei, alongside ambitions for a 5–10-trillion-parameter LLM and expanded computing capacity—an incumbent-led competitive push in AI infrastructure and models.
- Andy launched from stealth as a B2B corporate entertainment network intended to centralize fragmented enterprise booking and supplier workflows across dining, sports, events, merchandising, and catering. At launch, it reported 60 large-company contracts, including Monday.com, Cloudflare, and Salesforce, alongside more than 1,600 hospitality groups and 93,000 venues. Andy says its case studies show 12–15% savings and forecasts about $400 million in spend routed through the platform by year-end; the latter is a projection, not reported realized volume.
- Founder Loit Sarma is described as having co-founded an ADP-backed HR venture and spent 2.5 years at ADP overseeing its money-movement and payments portfolio. Andy describes its product as a virtual entertainment concierge with business context and multi-actor workflows; AI agents and a voice agent support quick responses and updates, while the company says it retains human involvement and is not yet having agents negotiate on customers’ behalf. The company also says it retooled product, design, engineering, and go-to-market processes to be AI-native after models became more capable.
- Andy’s founder flagged money movement, tax jurisdictions, venue onboarding, and activating new cities and global coverage as scaling challenges.
- Boom Supersonic is extending its internally developed supersonic engine core into trailerized, behind-the-meter data-center power: founder Blake Scholl described 42 MW deployments and said the hot-running core does not require water.
- Scholl said the first vertically integrated engine core was being assembled for a test stand the following month; the factory had begun making parts, with stated targets of multiple gigawatts of annual capacity and adding 10+ GW to the grid over five years.
- Scholl claimed tens of gigawatts of demand and said the company would auction capacity, but the exchange provided no named customers or revenue figures. The founder described Boom as 12 years old, making this an adjacent infrastructure signal rather than a clear Seed–Series A opportunity.
- California gubernatorial candidate Steve Hilton proposed reducing the state’s regulations from a claimed 420,000-plus to fewer than 200,000 by the end of his first term . He also announced a plan for 10 new cities, pitching them as part of making California a place to build businesses again .
- Spencer Pratt floated a hypothetical AI/data-center redevelopment concept for Skid Row: place data centers there and seek private investment tied to infrastructure upgrades such as water pipes; he also suggested retraining fentanyl users as electricians and referenced a Meta program .
Fei-Fei Li, a Stanford researcher and professor, is World Labs’ cofounder and CEO. World Labs is preparing to release Atlas, its latest world model. Unlike language models that output words, its world models generate pixels and 3D/4D geometry; World Labs describes Atlas as first-of-its-kind for camera control and pixel consistency, combining image generation with reconstruction of detailed 3D structure.
World Labs says it is engaging users in creative, design, and robotics fields, including healthcare and pharmaceutical labs. For robotics, the company sees world models as a way to address bottlenecks in training and evaluation environments; it says simulation can help create realistic training and evaluation grounds for data-scarce pharmaceutical manufacturing and lab-science tasks.
- World Labs co-founder and CEO Fei-Fei Li is a Stanford professor and AI researcher. The company is developing Atlas, a world model intended to generate pixels with 3D structure, allow camera control, and combine image generation with intricate 3D reconstruction; Li says it outperforms specialized state-of-the-art models on these tasks, and product release is in progress.
- Li describes applications across creative and design work, robotics, healthcare, and pharmaceutical labs. She sees realistic simulated environments as a way to address robotics training and evaluation bottlenecks, including data scarcity in specialized laboratory and pharmaceutical settings.
- Li says world-model evaluation should be an ecosystem-level responsibility, with independent academic and public-sector benchmarking alongside industry participation.
- World Labs is preparing Atlas, a world model designed to generate visually consistent 3D worlds with precise camera control; Fei-Fei Li says it combines image generation with 3D-structure reconstruction and claims it outperforms specialized state-of-the-art models on these tasks.
- A potential application is robotics training and evaluation: World Labs’ models could create realistic simulated environments for developing robot tasks where real-world data is scarce, including pharmaceutical manufacturing and laboratory science. Li also names creative, design, healthcare, and robotics users as target areas.
- Li says World Labs evaluates Atlas internally for safety and capability, while arguing that world-model benchmarking needs broader ecosystem participation, including independent academic and public-sector bodies.
Sam Altman says startups are naturally good at quickly assembling empowered, capable teams around urgent priorities, while sustaining that organizational agility as a company grows is difficult and underexplored .
Jev is presented as a decision-making model rather than a text generator, with speed claimed at up to about 200× that of current chatbots; the proposed use is to handle simpler AI-agent decisions quickly, potentially reducing latency and token use while a heavier model handles harder decisions. Its described ingredients are a decision-focused architecture, parallel evaluation of candidate answers, and RLCD training intended to calibrate confidence so systems can trust Jev or fall back to a more capable model. The evidence and novelty are preliminary: the video notes there is no official research paper, only a high-level website description; related open research exists, the underlying idea is not new, and the presenter characterizes Jev as partly new and amplified by hype.
OpenAI’s stated API goal is to offer the best model at every price point and across modalities; Altman corrected “video” to “voice” in the modality list. He also invited developers to build products on the API. Ahead of DevDay, Altman said OpenAI had more to launch than it could handle, but shared no specific launch details.
A founder is building document-collection software for one-to-five-person small-business purpose lenders, including equipment finance, factoring, SBA, and private lending, but lacks industry access to validate the product assumptions. Of roughly 74 LinkedIn leads, about half were poor matches; 18 messages produced one decline and one reply from a solo originator without purchasing authority, and the founder still lacks feedback from someone who handles borrower paperwork.
- Rogo is targeting dealmakers in finance, using AI across transaction workflows such as preparing and reviewing data rooms, due diligence, deal materials, and updates to firms’ systems of record. Its founder sees private markets as a strong wedge because much of the coordination and transacting is still human-run, while public equities are more automated.
- The vertical-AI investment thesis is that durable products need more than model access: firms need domain-specific data and integrations, compliance and auditability, and workflow infrastructure that can progress from copilots to autonomous execution. The founder says promising verticals have enough complexity to support that work, and that teams need domain expertise and a willingness to repeatedly rebuild as models improve.
- Rogo’s founder said he had two years of finance experience and that the company assembled a team of more than 100 people with experience at investment banks or investment firms. He also described finance-AI distribution as trust- and people-intensive, with a substantial engineering and product burden; individual productivity gains have not yet necessarily translated into measurable firm-level gains such as more transactions.
Sam Altman said GPT-6 Sol and Luna improve on their 5.6-family predecessors in intelligence, alignment, work output, coding, and computer use, while costing half as much per token and less per task . He argued that per-task pricing is the meaningful comparison and said he did not think any market offering was competitive on that basis; he also said lower-cost access should let people use much more AI .
OpenAI introduced GPT-6 Sol and Luna as faster, more affordable models that carry much of GPT-6 Astra’s strengths for work at scale. It says improved caching and inference efficiency enabled API prices 50% below GPT-5.6 promotional pricing.
Sam Altman said GPT-6 Sol and Luna improve on the 5.6 family in intelligence, alignment, work output, coding, and computer use; he also said they cost half as much per token and even less per task.
Garry Tan calls for teaching people to prompt and use AI across their pursuits, alongside greater appetite to solve problems for themselves and others—a broad AI-adoption and skills signal. The linked discussion advocates rethinking education for the AI era, including designing problems students need AI to solve rather than evaluating only unaided work.