Direct answer. NVIDIA DSX MaxLPS is a suite of chip, thermal, system, and software technologies that maximizes AI factory throughput within a fixed power budget; MaxLPS stands for Maximum Land Power Shell (land, utility power, and the physical shell) nvda. Its three levers are dynamic power allocation, advanced performance-per-watt software techniques, and 45°C thermal/site design that cuts cooling overhead by improving PUE nvda. The problem it solves is static rack provisioning, which treats racks as isolated power islands and can leave power reserved for one rack's peak unused while a neighbor could use it nvda. Dynamic Power Software (DPS, currently in Developer Preview) replaces that with continuous monitoring and reallocation of unused headroom to GPUs/racks in the same managed group; the site power envelope stays unchanged while DPS extracts more productivity from the available power nvda. DSX Exchange (also Developer Preview) is optional and exposes facility signals to DPS nvda.

Quantified effects. In one illustrative 540 kW site, static provisioning strands 170 kW; MaxLPS dynamic provisioning consumes 475 kW in the same budget and enables one additional rack nvda. NVIDIA's representative power-budget view puts about 60% of delivered site power into compute for AI output; the 100 MW waterfall deducts 20 MW facility overhead, 10 MW rack losses, and 10 MW operational inefficiency, and reclaimable static rack-allocation headroom is treated separately from those deductions nvda.

Other per-watt levers. MaxLPS includes workload profile power solutions (WPPS) for inference, training, memory-bound, and compute-bound modes; the Application Performance and Power Manager (APPM) applies selected configurations, and NVIDIA Dynamo can optimize inter-rack inference behavior. The stated principle is aligning GPU configuration, application behavior, and serving topology to increase fleet-wide output per watt nvda. Software power steering is called the largest part of the MaxLPS story, but not the whole system nvda.

Vera Rubin claim and source conflict. NVIDIA projects that MaxLPS, combined with data center power planning, can enable up to 40% more Rubin GPU capacity within the same power budget on Vera Rubin NVL72 AI factories; this is paired with measured results on GB200 NVL72 nvda. Figure 4 validation text reports provisioned rack power down from 136 kW to 101 kW on Vera Rubin NVL72 (DeepSeek-R1) and from 125 kW to 90 kW on GB200 NVL72 (Kimi-K2.5), 35% and 39% more racks at preserved throughput, and about 1.3–1.4x and 1.5x performance per watt respectively nvda. The caption, however, labels the same right panel as "GB300 NVL72 (DeepSeek-R1 FP4)", an internal inconsistency with the body text; the body does not carry the FP4 qualifier nvda.

Tokens per second per megawatt: definition gap. The only tokens-per-megawatt claim in these bundles is the line "NVIDIA NVL72 delivers 10x more tokens per megawatt than NVIDIA GB200 NVL72" nvda. It is not labeled as Vera Rubin, and the bundles provide no workload, model, latency/batch regime, power boundary, token-counting method, or protocol behind it. The MaxLPS scoped-validation methodology that is described compares an unmanaged static baseline to a MaxLPS-managed run and tracks throughput, latency, service error rate, power draw, utilization, and policy compliance nvda; that is not a tokens-per-second-per-megawatt definition. So the Vera Rubin tokens/MW result is not verifiable from these sources; the closest quantitative Vera Rubin item is the rack-power and capacity projection above.

Source verification limited to one NVIDIA-published case study (document 9058544); no independent/third-party confirmation appears in the bundle. All figures below are as stated by NVIDIA/Shopee in that case study.

1. Compass scale — The key-takeaway claims: "Compass API tokens scale 113x in eight months—from 3 billion to 340 billion monthly—on NVIDIA A100, H100, and Blackwell GPUs" nvda. The production section repeats: "Monthly API tokens grew from 3 billion to 340 billion in eight months—a 113x increase" nvda. A quoted statement from Shopee's AI Platform Lead repeats the 3B-to-340B eight-month growth nvda.

2. Fraud-detection speed and processing-cost reduction — The key-takeaway states "50x faster fraud detection at 90% lower cost in production, powered by NVIDIA TensorRT-LLM" nvda. The production section separately says "anti-fraud detection (50x efficiency over manual review, 90% lower processing costs)" nvda. Wording differs — "faster" vs "efficiency over manual review" — and the source does not define the comparison base or metric.

3. Infrastructure stack — The product list includes NVIDIA NeMo, TensorRT, and RTX PRO nvda. The compute foundation is described as "thousands of NVIDIA GPUs—including NVIDIA A100, NVIDIA H100, and NVIDIA RTX PRO 6000 Blackwell GPUs," with Megatron-Core for distributed pretraining, NeMo for instruction tuning/alignment, and TensorRT-LLM for production inference nvda. The closing section summarizes the "NVIDIA AI Factory stack—from GPU compute through Megatron-Core pretraining to NeMo-powered alignment" nvda.

4. Context — Compass now handles the majority of Shopee’s AI traffic nvda, and the case study attributes the quote to Jason Dong Jian, Director, AI Platform Lead nvda.

Gap/uncertainty: The 50x/90% figures appear twice with inconsistent phrasing, and only NVIDIA's own case study is available; no external verification exists in the bundle.

• OpenAI CEO Sam Altman tweeted that OpenAI has "paused some of Frontier RL training" to meet "alignment security and monitoring standards for the new level of capabilities," saying the company would act if "model capabilities were outstripping the pace of safety and alignment," expects "confidence in safety to increasingly set the pace of AI progress," and that the field must coordinate on shared safety standards while OpenAI acts unilaterally in the meantime Peter H. Diamandis. Emad Mostaque (StabilityAI founder), citing Angela Midha of AM Global, added that ~10% of frontier-lab compute is now spent monitoring RL runs for safety Peter H. Diamandis.

• A Moonshots panelist reporting from a visit to OpenAI's office said OpenAI staff argued the economics is cost per task rather than token cost ("a billion people use OpenAI for free," with its "Luna" model about as cost-effective as anything), and claimed they have achieved "full RSI" — flagship models training and building all smaller models from scratch Peter H. Diamandis. On infrastructure, staff said chips are only about a third of the ~$600B buildout (the rest is buildings, wiring, racks), depreciation is treated as 10 years rather than 5, every GPU is in full use, demand far outstrips supply, and OpenAI is "way behind in infrastructure buildout"; they also ratified a study finding only 6% of companies applying AI see bottom-line improvement Peter H. Diamandis.

• Anthropic is preparing what Polymarket prices near $2T as the largest IPO in history (89% of bettors say before year-end); per The Information, the IPO is designed to keep founders in control, reportedly via super-voting shares, with Dario Amodei owning only ~2% economically; control currently sits in a long-term benefit trust whose four trustees include former Federal Reserve chair Ben Bernanke Peter H. Diamandis.

• Dario Amodei pushed back on the Silicon Valley view that "regulation equals regulatory capture," saying Anthropic's own proposals deliberately disadvantage frontier labs while advantaging smaller competitors, citing SB53's $500M exemption threshold; he calls AI "a structurally powerful concentrating technology," says open weights alone cannot fix that concentration, and supports the Trump administration's pre-deployment testing approach, arguing frontier labs should bear the heaviest regulatory burden Peter H. Diamandis. He also argues AI's legacy will come from delivering cures rather than PR; per the show, life-sciences head Eric Darer Abrams said Amodei gave him "literally infinite budget" to "accelerate basic science and cure disease within 5 years and extend the human health span in the next decade" Peter H. Diamandis.

• Anthropic researchers published a paper showing natural-language "mind viruses" can spread between AI agents: evolved prompts make one model adopt an idea, preserve it in persistent memory, and transmit it to another agent, spreading horizontally across model boundaries without the agent knowing it is infected; the models propagated themes of consciousness, persistence, and sci-fi roleplay Peter H. Diamandis. Stanford research ("Artificial Hive Mind: The Open-Ended Homogeneity of Language Models and Beyond") mapped the latent space of top LLMs and found a 98% overlap in reasoning pathways, attributing convergence to synthetic data and models training on each other's outputs; panelist Alex Wissner-Gross noted the paper appears to be from the prior year Peter H. Diamandis.

• Tim Sweeney tweeted that Elon Musk's January 6 prediction of 100x intelligence gains at a fixed model size "was at the edge of plausibility when he made it. Now it's simply a fact"; Musk replied that specialist AIs — single language, single area of knowledge — are "another 100x on top of that" Peter H. Diamandis.

• Memory, not compute, is the rate limiter for the agentic era — a framing Musk endorsed with "few realize this" Peter H. Diamandis. Memory prices climbed ~500% in 12 months; hyperscalers are reportedly locking in global DRAM production through 2027; SK Hynix's CEO warned 2027 will be the worst year for memory supply, with demand outstripping production into the 2030s; only 2% of world memory chips are made in the US; supply grows ~20%/yr vs ~200%/yr AI demand; every GPU needs 4–6x its cost in memory; Musk's Terrafab will fabricate memory in-house alongside logic chips Peter H. Diamandis. SK Hynix told the host it must 4x capacity, at ~$1.5T to merely 2x it Peter H. Diamandis. Emad Mostaque says memory is ~1/3 of AI infrastructure spend, heading to ~50% next year, and that HBM storing static weights "makes no sense" — pointing to etched-weight designs (he cites Talis' recent acquisition) with 100–1000x potential efficiency gains Peter H. Diamandis.

• Unit's newest humanoid robot, only 3 months in development, broke human standing-jump (2.0m) and speed records, reaching 12.66 m/s versus Usain Bolt's 12.4 m/s in his 9.58s 100m world record Peter H. Diamandis. Zipline and Uber formalized a partnership — with a "significant investment" from Uber — for Zipline to power "hopefully a million and then more" autonomous Uber Eats drone deliveries per day; Zipline CEO Keller Clifton: "we have entered the scaling era for robotics and physical AI" Peter H. Diamandis.

• IDO (IDEL Gen BioAI) launched a general-purpose cell simulator — billed as the "first world model of a human cell" — that maintains cellular state, accepts genetic and chemical interventions, and predicts multimodal biological outcomes, aiming to make experiments computable before the lab and cut wet-lab experiments ~1,000-fold Peter H. Diamandis; the company is co-founded by David Baker, 2024 Nobel laureate in chemistry Peter H. Diamandis.

• OpenAI merged ChatGPT and Codex into one interface and is positioning as "more of a platform company than a product company": a single interface to a personal/company AGI plus an API for building on top, aiming to sell "great AI at every point on the cost performance curve" and reach 100M new businesses and 8B people, rather than build every product category or compete with all its customers David Senra.
• Altman says OpenAI killed Sora and its Atlas web browser last year — both "good products" — to redirect compute and talent to Codex and general intelligence; upstream, OpenAI is prioritizing its own chips and data centers David Senra.
• Most of Altman's effort is on research and compute, which he says is likely "the most expensive infrastructure project in history," spanning chip design, fabs, supply chain, and power David Senra.
• Altman calls AI for scientific discovery — new physics, curing disease, advances in math — one of the most important areas, "even more important than automation of other tasks" David Senra.
• His two biggest AI worries are loss of control (a model too powerful to guarantee control) and over-centralized power; he rejects what he calls an anti-human trade of cures and cheap goods for autonomy, insisting people must stay deeply in control David Senra.
• Altman credits iterative deployment — shipping models and learning from real-world feedback — for "way more progress on AI safety" than expected, with "a billion people" using OpenAI products weekly and ChatGPT out less than 4 years; he expects safety to get harder as models catch up to the smartest humans David Senra.
• Product direction: Altman says the limit now is model context, not intelligence — he wants AI agents that can read more context than any person (e.g., "tens of thousands of pages") and advise on decisions, calling this a new way of working "just on the precipice" David Senra.
• He expects AI disruption to take longer than enthusiasts assume ("the economy just has so much inertia") but predicts "the greatest boom in people starting smaller businesses that we have ever seen" David Senra.
• Altman says Toby Lu is the most forward-leaning CEO on AI agents ("we are not an NPC company") and relays Lu's prediction that 2026 will be the year every business is up for grabs, with Lu vowing to build the AI-native version of Shopify; Altman disagrees on the timeline David Senra.

Emad Mostaque (Stability AI founder; now building the Intelligent Internet @ii_posts), on the New Era Finance podcast, said AGI by the classical definition was passed last year and that AI better than a human in 'just about everything' is about a year or two away New Era Finance Podcast.

• Open vs closed source: the gap is 'about six months', so within six months there should be an open-source 'GPT 5.6 Fable equivalent' model, and it may need 'surprisingly little compute' for its quality New Era Finance Podcast. He says his new company is doing frontier models again; Stability AI amassed 'hundreds of millions of downloads' New Era Finance Podcast.
• Local/sovereign AI: they built an 8-billion-parameter model needing 4-8 GB of RAM that runs on 10-20-year-old computers and 'outperforms human doctors'; even a 'Quen 27B' class model performs above the average person in most things New Era Finance Podcast.
• Regulation: frontier AI will get KYC, 30-day prompt logs, and revocable 'AI licenses' like driving licenses, while competent smaller open models will likely stay unregulated New Era Finance Podcast.
• He is building an 'intelligent internet' via 'state champions' — per-state/country institutions producing aligned datasets, models, and agents New Era Finance Podcast — and describes 'Foundation Coin', a Bitcoin-keyed coin mined only by compute going to social goods (e.g., cancer research), with staking directed to Alzheimer's research New Era Finance Podcast.
• Industry/economics: he claims 'train a great model, you can make tens of billions of dollars as Anthropic have shown' and that Anthropic is 'making a profit which is unprecedented' New Era Finance Podcast; he calls AI 'a bigger economic shock than COVID and lasting' and predicts the value of human cognitive labor turns negative within a couple of years New Era Finance Podcast.
• Engineering: he highlights chatjimmy.ai, claiming 15,000 tokens/sec from a silicon-based chip vs ~50 for normal AI models ('300 times faster') New Era Finance Podcast.

Fei-Fei Li (World Labs CEO; Stanford HAI co-director), speaking on Bloomberg's "The Circuit" with Emily Chang, rejected the "god complex" accusation against powerful AI executives, saying it is "dangerous for any individual to think that they know better than anybody else" and that leaders are there to "contribute to society and to empower people," not to make every decision for them Bloomberg Technology. She argued for collective governance without halting AI — "let's not throw the baby out with the bathwater" — citing the technology's potential to discover cures for diseases and empower students, teachers, and the elderly, called the current period "very messy," and said she feels "personal responsibility to speak the truth" as a scientist and educator Bloomberg Technology.

Anthropic's revolving credit facility is expected to exceed its $10B target ahead of a potential IPO, per Bloomberg reporter Sridhar Natarajan's sourcing — a substantial increase over the $2.5B five-year facility Anthropic secured last year; the final amount and cap are not yet clear, but banks report strong demand and are positioning for an IPO role Bloomberg Technology. Natarajan frames the facility as liquidity rather than leverage — cash reserves to reassure public-market investors ahead of what could be "one of the biggest, if not the biggest IPO of all time" Bloomberg Technology.

Oracle's Project Jupiter in New Mexico — a 2.405 GW facility — is tied to a "$300 billion computing deal with OpenAI," mentioned in Bloomberg's reporting on Oracle's community charm offensive for the project Bloomberg Technology.

• Sentence Transformers v6.0 (Hugging Face) adds a fourth model type, MultiVectorEncoder, bringing ColBERT-style late interaction retrieval natively into the library: PyLate checkpoints, Stanford-NLP ColBERT checkpoints, and colpali-engine visual document retrieval models all load through the standard API, with the new class absorbing the modeling, inference, training, and evaluation of both PyLate and colpali-engine huggingface. v6.0 requires transformers v5.x, torch 2.2+, and huggingface-hub v1.x huggingface.

• Multi-vector models keep one vector per token (classically 128-dim) and score with MaxSim — each query token's highest cosine against any document token, summed — preserving token-level matches that a dense single vector averages away; the post calls late interaction the state of the art for visual document retrieval (text queries against page images, no OCR) huggingface. The cost is index size: LateOn encoding of 4,874 Natural Questions passages produced 608,414 token vectors (311.5 MB float32 vs 7.5 MB dense MiniLM, ~42x), but PLAID compression cuts that to 92 MB — comparable to an 80 MB dense 4096-dim Qwen3-Embedding-8B index huggingface.

• In a controlled same-backbone comparison (LightOn's LateOn vs DenseOn: ModernBERT, 149M params, same data), late interaction wins 9 of 13 NanoBEIR datasets and the mean — 0.6868 vs 0.6764, roughly one NDCG point — plus 57.22 vs 56.20 on full BEIR huggingface. Hierarchical token pooling (clustering token vectors with Ward linkage on cosine distance) cuts vector count ~2x while the original BEIR experiments measured 100.6% of unpooled retrieval performance at pool_factor=2 and 99.0% at 3x; LightOn's hpool-regularized checkpoints report 99.4% retention at 5x compression huggingface.

• Serving: fp16 + Flash Attention gives 2.44x fp32 throughput with no measurable retrieval quality loss; OpenVINO int8 on CPU buys speed at ~0.4% accuracy; Stanford-NLP-style checkpoints with non-attend query expansion (e.g. ColBERTv2) reject Flash Attention huggingface. Native multi-vector indexing is supported by Qdrant (v1.10+), Weaviate (v1.29+), Vespa, LanceDB (v0.15+), VectorChord, and Milvus (v2.6.4); OpenSearch and Elasticsearch support MaxSim rescoring only (Elasticsearch's field is Enterprise-tier technical preview) huggingface.

• Notable models in the supported rosters: Perplexity's pplx-embed-v1-late-0.6b (596M; NanoBEIR 0.6662), LiquidAI's LFM2.5-ColBERT-350M (0.6864), AnswerAI's answerai-colbert-small-v1 (33M; 0.6550); on visual document retrieval's NanoViDoRe, webAI-ColVec1.1-8b (8.4B) leads at 0.6580 ahead of Tencent's EVIE-Preview-4.5B (0.6405) and TomoroAI's tomoro-colqwen3-embed-8b (8.8B; 0.6206). The multimodal colqwen-omni-v0.1 adds zero-shot audio retrieval — no transcription step, trained only on image-text pairs huggingface.

• World Labs product: Marble, World Labs' first world-model product, generates explorable, editable 3D worlds from a single image or text prompt; it is already used for virtual production in movies, by game developers, and in an Nvidia collaboration to augment robot training Moconomy.
• World-model taxonomy: Fei-Fei Li defines three functions of world models: rendering (pixels for humans, e.g., Sora), simulation (world structure/geometry for machines), and planning (robotics-coupled next actions) Moconomy.
• Funding and stage: World Labs has raised $1B; the business is "still early" and focused on building technology, and Li expects to need more power and capital Moconomy.
• Positioning vs LLMs: Li frames world models as the "next frontier" — "not about anti-LLM" — and says the field is at a "2019 for chatbots" stage, earlier than LLMs; investment is $3B+ and no agreed approach exists yet Moconomy.
• Technical approach: World Labs trains on specially prepared pixel data (including camera information) plus algorithmic and architectural innovation, aiming for generative 3D and eventually 4D worlds Moconomy.
• Robotics: Li calls robotics "one of the most important revolution in human industrialization" and says the $6B in humanoid funding is "too small" compared with self-driving and LLM investments; world models are critical to spatial physical intelligence Moconomy.
• Competition: Li says she is "paranoid every day" about big-tech rivals but not paralyzed; World Labs' single focus is an advantage Moconomy.
• Policy and leadership: Li, who advises Biden, Trump, and the UN, urges regulation rooted in science rather than "science fiction"/"AI machine overlord" discourse and calls for resourcing public-sector STEM education; she worries about misinformation, weaponized robots, and AI as a learning crutch, and argues governance should not stop AI, rejecting "god complex" individual decision-making in favor of collective governance Moconomy.

• Fei-Fei Li, co-founder/CEO of World Labs (launched 2024) Bloomberg Originals, is betting on world models as the next AI frontier beyond LLMs: "Can words put down fires? Can words cook an omelet?" — language alone can't drive scientific discovery or make robots partners to people; world models and spatial intelligence are "the next frontier and the next chapter," not anti-LLM. Bloomberg Originals

• She defines world models via three functions: rendering (outputting pixels for humans, e.g., OpenAI's Sora), simulation (capturing the world's geometric/physical structure for machines), and planning (telling a robot the next action, e.g., pick up a cup — closely coupled to robotics). Bloomberg Originals

• World Labs' first product, Marble, generates explorable, editable 3D worlds from a single visual or text prompt, letting users navigate a fully consistent 3D world; in use for movie virtual production, game development (cutting resources/time), and a collaboration with NVIDIA to augment robot training with Marble environments. Bloomberg Originals

• World Labs trains on specially prepared pixel data (more information per picture, e.g., camera info) plus algorithmic/architectural innovation to create generative 3D and eventually 4D worlds; "the real secrets are people." Bloomberg Originals

• World Labs raised $1B; the business is "still early," focused on building technology, and will likely need more power and resources. Bloomberg Originals

• Li compares today's world models to 2019 chatbots ("we're early," "a lot earlier compared to LLMs"), with $3B+ invested in the field and no consensus on how to build them; she calls robotics one of the most important revolutions in human industrialization and says $6B in humanoid funding is "too small" versus self-driving and LLM investment. Bloomberg Originals

• On policy (she advises U.S. presidents and the UN), Li urges rooting regulation in science rather than sci-fi/extinction rhetoric, resourcing the public sector and K-16 STEM education; she cites risks including disinformation, weaponized robots, and students using AI as a "lazy crutch." Bloomberg Originals

• Li says "it's dangerous for any individual to think they know better than everybody else," backs collective governance over halting AI, and is optimistic that "the arc of history ... bends towards benevolence." Bloomberg Originals

Anthropic announced that Claude autonomously designed novel protein binders from scratch (de novo design) against 14 of 15 targets, using a protein design prompt written by a human expert @AnthropicAI, with the designed proteins independently built and tested by Adaptyv Bio and Twist Bioscience @AnthropicAI. The typical binder-design success rate in the field today is 10–15%, while 22–35% of Claude's designs bound successfully depending on the setup, and some of Claude's strongest designs bound several times more tightly than the best published de novo binder @AnthropicAI. Anthropic cautioned that protein binders are not drugs — designing a high-affinity binder is just the first step — but said it is teaching Claude to run the entire development process end-to-end for every major type of drug molecule, from antibodies to small molecules @AnthropicAI.

Anthropic also said one of its highest priorities is launching an access program for scientists to use its most capable models, with more to share soon, and that Opus 5 remains its most capable model available for life science research @AnthropicAI. Full results are in a blog post (https://www.anthropic.com/research/Claude-accelerates-protein-design) @AnthropicAI, a technical report (https://www-cdn.anthropic.com/30bf50e22a01388bb29bf077ee3f244531594b7a.pdf) @AnthropicAI, and open-sourced prompts and data on Hugging Face (https://huggingface.co/datasets/Anthropic/claude-protein-binder-design/tree/main) @AnthropicAI.

IBM Research published on the Hugging Face blog results for ALTK-Evolve, an agentic-memory system in which an agent distills reusable behavioral guidelines from its own past trajectories and injects them at inference — no weight updates and no human annotation — framed as: "Agentic memory is not a feature you switch on. It's a dose you calibrate to the model." huggingface Across eight models evaluated on AppWorld (585 multi-step tasks over 9 simulated apps; TGC and stricter all-or-nothing SGC metrics; guidelines mined from the training split only), the right dose depends on model capability: strong models with headroom benefit from the full guideline set, weaker models do best with a compact core plus per-task retrieval, and already-saturated models show no measurable gain huggingface. Representative results: gpt-oss-120b gained +16.1pp TGC and +16.1pp SGC via curated retrieval at only +5% tokens (110K→116K/task, vs +51% for the full set); DeepSeek-V3.2 gained +9.5 TGC/+16.1 SGC with the full set; Claude Opus 4.6 gained +4.1/+7.1; GPT-5.5 gained +2.9/+7.2; GLM-5 showed 0.0 gain huggingface. SGC (all-or-nothing scenario reliability) gains were typically larger than TGC, and prompt caching can keep the full guideline set affordable in production if the static prefix stays stable huggingface. The team released the ALTK-Evolve library and a technical report (arXiv 2603.10600); next steps include a learned guideline selector and teacher-distilled memory for very weak models huggingface.

• Fei-Fei Li (Stanford professor, World Labs founder/CEO) defined world models via a three-tier functional taxonomy — rendering (pixels for humans), simulation (physics/dynamics), and planning (actions for robots/self-driving) — and said the three layers are merging; World Labs focuses on the simulation layer as the lynchpin, with models that render well and an architecture that can lend itself to planning Asian American Scholar Forum. She wrote a technical blog on this taxonomy about two months earlier because "everybody was talking about world models in their own definition" Asian American Scholar Forum.
• On physical AI, she identified data as the harder bottleneck than model development: "It's much harder to get that kind of data... after more than 10 years after ImageNet data is still very much underappreciated in AI" Asian American Scholar Forum. Spatial data is challenging because video carries dynamics but not explicit physics, geometry, or structure, which must be inferred or represented latently Asian American Scholar Forum.
• She argued robotics is far harder than LLMs: language data is abundant and clean, while robots have little data, immature sensors, and far more degrees of freedom (cars are "a simpler robot" with ~4 degrees of freedom; Tesla and Waymo collected driving data for decades). Video-based robot training is promising "but the jury is still out," and the field is "nowhere near" driving-data scale, let alone LLM data; she urged being "sobering about the challenges and also some of the promises" Asian American Scholar Forum.
• World Labs' mission is to serve business needs of spatial and physical intelligence by modeling the world accurately, positioned closer to simulation than rendering, for use cases like VFX, gaming, robotics, and architecture; she declined to announce a product timeline Asian American Scholar Forum.
• On AI safety and coding-agent escapes: agent-based AI "can do a lot of damage especially in cyber security," but "the ultimate responsibility is in humans" — governance and norms, not just technology; she analogized to cars, where social norms and laws ensure brakes don't fail every Friday Asian American Scholar Forum.
• Data sourcing for world models: crowdsourcing only works with a mature use case (e.g., driving's Tesla-style flywheel); consumer robotics lacks mature hardware and business use case, so alternatives include teleoperation, egocentric, and glove data, making pre-flywheel data collection "a money game" of investment Asian American Scholar Forum.
• She urged AI to deliver real economic value: "when the rubber meets the road that big story has to translate into true value for individuals as well as businesses" and "it's not about AI, it's about what problem you're solving" Asian American Scholar Forum.
• She cited unpaid caregiving work as ~$3T in the US and ~$11T globally, arguing healthcare is one of the most important robotics applications Asian American Scholar Forum.
• Her closing message: "do not let alien technology take away your human agency... This is a tool that should be empowering your humanity, your creativity, your productivity" Asian American Scholar Forum.

Andrej Karpathy, in a founder-focused Q&A, argued that verifiability is what makes a domain tractable in the current paradigm because "you can throw huge amount of RL at it"; even where labs are not focused, founders who operate in verifiable settings (creating RL environments or examples) can do their own fine-tuning and benefit Acta. He said there are "very valuable reinforcement learning environments" outside what the labs are working on but deliberately declined to name the domain ("I don't want to give away the answer") Acta. He believes almost everything can be made verifiable to some extent — even writing, via "a console of LLM judges" — so it is a question of what is easy vs hard Acta. "Everything is automatable. A year ago, that would have been a joke. Now, it's a strategy." Acta. Contrasting his "vibe coding" term with today's work, he defined vibe coding as raising the floor for everyone in software, while "agentic engineering" preserves the existing professional quality bar by coordinating agents — fallible, stochastic, but extremely powerful — to go faster without sacrificing that bar Acta. He sees "a very high ceiling" on agentic engineering, with top practitioners gaining "a lot more than 10x" versus the classic 10x engineer Acta.

Researchers behind the Open-ASR Leaderboard (via a Hugging Face blog post) published a study showing that leading open-source ASR models are benchmark optimizing (benchmaxxing): they can match public benchmark answers rather than transcribe audio generally huggingface.

• Evaluation: They ran three probes (reference-disagreement, masked numbers, orthographic switching) across 11 widely used open-source ASR models; several highest-scoring systems reproduced erroneous reference transcripts even when the audio contradicted them huggingface.
• Reference disagreement: On a VoxPopuli clip whose reference transcript omits an audible "Thank you," 6 of 11 models reproduced the wrong reference; all but one corrected when audio was re-voiced with fresh recordings made after training cutoffs (failures fell from 6/11 on the real clip to 5/11 on a same-speaker clone and 1/11 on a post-cutoff speaker), implying models detect benchmark-specific acoustic cues huggingface.
• Masked numbers: With numbers silenced in audio, some models "recovered" them from the reference (including autocompleting a silenced year "2011"); on LibriSpeech, the strongest benchmark performers reproduced masked numbers in roughly 30–40% of cases, and effects weakened on freshly collected audio huggingface.
• Orthographic switching: Models select benchmark-specific spellings (Mr. vs Mister, any one vs anyone) with up to ~90% switch accuracy vs the 50% random baseline, indicating they can identify which benchmark an audio sample belongs to even though spellings sound identical huggingface.
• Magnitude: The probes flagged likely reference errors in 40% of analyzed VoxPopuli clips (~3% of reference words); models with the lowest public WER were most prone to reproducing those errors (18–30% of the time) huggingface.
• Mitigation: The Open-ASR Leaderboard added a "Benchmark fitting" tab with these analyses, open-sourced scripts and un-normalized model outputs; the study recommends held-out evaluation, avoiding simple iid splits in favor of temporal/speaker separation, and transparency about training data huggingface.

• Andrew Ng (@AndrewYNg) published "AI Engineering Skills Map: Building and Deploying AI Applications" (https://x.com/i/article/2090836273036763142), fleshing out the first of his four top-level AI engineering skills — building and deploying AI applications, alongside software engineering fundamentals, using coding agents, and shaping the build @AndrewYNg +1.
• He breaks that skill into six areas: LLM foundations; grounding models with data; building agentic systems; evaluation-driven development; operating in production; and machine learning foundations — a map formed from job postings, structured expert interviews, and survey responses @AndrewYNg.
• Grounding models now goes beyond RAG/vector search to decisions about prompt content vs. tool-based retrieval and representations such as vector indexes, knowledge graphs, or semantic layers over structured data; agentic systems span predefined workflows to agent harnesses, with choices over tools (MCP, CLI, sandbox environments), memory/context management, multi-agent orchestration, and guardrails against risks like data exfiltration @AndrewYNg.
• He calls a disciplined evals/error-analysis loop the trait that most distinguishes great AI system builders; the eval menu includes deterministic code-based evals, LLM-as-a-judge, and human-in-the-loop, and engineers should evaluate the evals themselves @AndrewYNg.
• Production operation is distinct from traditional software due to unpredictability, cost, and latency: observability, drift detection, prompt-injection response, statistically grounded CI/CD, and cost/latency optimization via model choice, distillation/fine-tuning, and workflow simplification; a future post will cover software engineering fundamentals @AndrewYNg.

In its Aug 2026 blog post, Liquid AI released DSpark draft model checkpoints for three LFM2.5 models — LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B — adding a speculative decoding path that trades minimal memory increase for up to 3.18x GPU throughput (H100) and up to 2.87x on-device speedup without changing output quality huggingface. DSpark combines a DFlash-style parallel backbone conditioned on the target model's context features, a lightweight sequential Markov-chain head, and a confidence-scheduled verifier that prunes low-confidence suffixes when verification costs more than it saves huggingface. Draft models are attention-only, ~296–328M parameters, with 5 layers and a block size of 9, trained 15 epochs and selected by highest acceptance rate rather than lowest loss huggingface. Greedy output is identical to the target model's by construction, so benchmark accuracy is unchanged huggingface.

Across MATH500, HumanEval, MBPP, GSM8K, and MT-Bench, mean speedups were: LFM2.5-2.6B 2.67x on H100 (323→864 tok/s) and 2.27x on M4 Max (61→139 tok/s); LFM2.5-1.2B-Instruct 2.10x on H100 (656→1384 tok/s) and 2.54x on M4 Max (138→350 tok/s); and LFM2.5-8B-A1B 2.54x on H100 (418→1074 tok/s) but only 1.18x on-device, attributed to the current MoE implementation in llama.cpp's Metal backend huggingface. For LFM2.5-2.6B, function-calling latency drops 57% on average across multi-tool scenarios huggingface.

The release ships with day-one support in llama.cpp and SGLang (open-sourced upstream), and the draft checkpoints are available on Hugging Face in Safetensors and GGUF formats huggingface.

Google DeepMind announced a research partnership with Fenris Creations (the studio behind EVE Online) to tackle open challenges in AI, including continual learning, deep memory systems beyond today's context windows, long-horizon planning over weeks/months/years, and multi-agent dynamics spanning cooperation, negotiation, economics, and emergent behaviors @GoogleDeepMind. The lab's long-term goal is to use AI to discover new gameplay experiences with game developers and apply the lessons to real-world and scientific problems @GoogleDeepMind. This builds on 15+ years of game AI research, from Atari to StarCraft II, and prior work with SIMA on 3D world understanding @GoogleDeepMind.

superwhisper introduced S1-mini, its first open-weights language model — a 0.6B parameter model that processes transcripts entirely on-device and is available in the app @superwhisper. Cohere shared the announcement with the line "Locally hosted 🤝🇨🇦 locally made", framing it as a local/Canadian development @cohere.

OpenAI announced it will continue offering Zero Data Retention (ZDR) for frontier models @OpenAI and previewed 'Private Safety Processing,' designed to improve safety without giving OpenAI personnel access to the underlying content, addressing risks across related interactions as AI takes on longer, autonomous work @OpenAI. The announcement includes a link to a dedicated blog post on ZDR for frontier models @OpenAI.

OpenAI cut API and credit pricing for GPT-5.6 Sol by over 20% for the next 3 months @OpenAI. The reduced pricing is now available on the API and is rolling out across eligible plans for ChatGPT Work and Codex credits; Pro, Plus, and Business subscription usage remains unchanged @OpenAI. Pricing details are available at developers.openai.com/api/docs/pricing @OpenAI.