April 20 AM: Robots master cluttered workshops & looped models beat size limits & PM cull begins & OSS agents ready
DeepMind claims robots now count tools in clutter and read gauges to sub-tick accuracy.
Robots Master Cluttered Workshops
DeepMind claims its latest model lets robots pinpoint tools in messy workshops, verify tasks, and read gauges accurately, moving industrial automation closer to reality.
The releases describe a model that fuses live camera streams, self-corrects distorted patrol images via generated code, and decides whether to retry tasks. [1] Think of it like giving a robot the spatial common sense a warehouse worker has after one week on the job. For founders building physical products or services, this could slash deployment costs and timelines for robots in real factories instead of controlled labs. Yet the provided counter-claims note these are promotional quotes without accuracy rates, failure modes, or tests under variable lighting and novel occlusions. Performance likely drops in true clutter. The emerging view is promising direction but not yet proven deployment-ready. Connects to looped models thread as both target efficiency in practical settings. [2]
“The claim is based on a vague marketing quote without quantitative metrics (e.g., accuracy rates, test conditions, or failure modes).”— Counter-analysis [2]
Sources (2)
- X post 2026-04-20 — Google DeepMind“Gemini Robotics-ER 1.6 upgrades visual and spatial understanding, enabling robots to pinpoint objects in cluttered environments, fuse multi-view camera streams for task completion verification, and read analog instruments with sub-tick accuracy.”
- Gemini Robotics counter-claim — Counter-analysis“The claim is based on a vague marketing quote without quantitative metrics (e.g., accuracy rates, test conditions, or failure modes).”
Looped Models Beat Size Limits
A new architecture from Together AI trains looped transformers stably at higher learning rates and matches 1.3B performance from 770M parameters, but the theory behind it is contested.
Together AI positions Parcae as solving the instability that has plagued looped models, where the same layers run multiple times for deeper thinking under fixed compute. [1] The payoff is real on benchmarks and lower inference memory. Yet the tracked counter-claims are blunt: 'Treating looped recurrence purely as a discrete LTI system is a severe oversimplification that ignores nonlinear activations, normalization layers, and position-dependent computations.' [2] Another notes the claimed learning-rate divergence 'is likely highly dependent on specific implementation details.' The evidence shows empirical wins at 140M-1.3B scale, but the theoretical grounding remains contested. No clear winner on whether this becomes the default path for efficient models or another clever trick with hidden limits. For any builder paying inference bills, 40% fewer parameters at same quality changes the economics immediately. [3]
“370M model scores 20.00 on Core vs 17.46 (+14.5%) and shows 6.3% lower validation perplexity than prior looped models.”— Together AI [3]
Sources (3)
- X post on Parcae 2026-04-20 — Together AI“Parcae introduces stable looped architectures by modeling recurrence as a discrete LTI system and constraining the spectral radius below 1 with a learned negative diagonal matrix, enabling training at LR 1e-3 versus 4e-4 for unconstrained loops.”
- Parcae counter-claim entry — Counter-analysis“Treating looped recurrence purely as a discrete LTI system is a severe oversimplification that ignores nonlinear activations, normalization layers, and position-dependent computations; a learned negative diagonal matrix offers only coarse, approximat...”
- X post on Parcae 2026-04-20 — Together AI“370M model scores 20.00 on Core vs 17.46 (+14.5%) and shows 6.3% lower validation perplexity than prior looped models.”
PM Cull Begins
Lenny Rachitsky warns that AI will eliminate half of today's product manager roles in a chaotic two-year reinvention, with traditional skills becoming liabilities.
Rachitsky describes a renaissance where PMs escape information shuttling without authority and instead connect instincts straight to customer tests with fewer dependencies. [1] The dark side is brutal: 30,000 PM jobs cut, only 8,000 replaced by roles demanding new skills. Analogy: like when cloud computing obsoleted certain sysadmin jobs in 2012; the survivors became infrastructure coders. Founders should care because your product velocity in 2027 depends on whether you retain or hire people who treat AI as a core building tool rather than a feature request generator. No real counter in the data, which itself is notable. This thread stands apart from the model and robotics work as the human side of the same efficiency wave. [2]
“Product managers face unprecedented chaos over the next two years as AI disrupts traditional PM roles, with half unlikely to adapt and survive.”— Lenny Rachitsky [1]
Sources (2)
- X post on PM upheaval 2026-04-20 — Lenny Rachitsky“Product managers face unprecedented chaos over the next two years as AI disrupts traditional PM roles, with half unlikely to adapt and survive.”
- X post on PM upheaval 2026-04-20 — Lenny Rachitsky“Companies will shed 30,000 PMs and rehire only 8,000 in AI-centric roles, amid widespread 'smiling exhaustion' among top performers.”
OSS Agents Ready for Production
Open-source models now match closed ones on inference cost, while new tools give agents persistent user-specific memory and clean subagent communication.
Chase notes inference now dominates costs at companies like Lindy, making OSS viable. DeepAgents seeds each user with a personalized AGENTS.md that persists across conversations and supports validated structured returns from subagents. [1] In plain English, your agents can remember user preferences at scale without custom databases and talk to each other without hallucinated JSON. This is the Uber moment for agent deployment: suddenly the economics work without relying on one provider. Founders running AI features should audit their stack now; the 2-5x cost drop changes unit economics. No major counters surfaced here, making the convergence toward open-source production notable. [2]
“Open-source models like GLM-5.1 have rapidly progressed to match frontier closed-source performance for most production use cases, particularly in inference efficiency.”— Harrison Chase [1]
Sources (2)
- X post on OSS parity 2026-04-20 — Harrison Chase“Open-source models like GLM-5.1 have rapidly progressed to match frontier closed-source performance for most production use cases, particularly in inference efficiency.”
- X post on DeepAgents 2026-04-20 — Harrison Chase“Deepagents Deploy now supports user-scoped memory via a user/ directory, providing each user a personalized writable AGENTS.md file.”
The open question: If looped models, robot perception, and agent memory all scale faster than expected, do we retrain our teams and infrastructure faster than the economic disruption hits?
- Google DeepMind — X post 2026-04-20
- Together AI — X post on Parcae 2026-04-20
- Counter-analysis — Parcae counter-claim entry
- Lenny Rachitsky — X post on PM upheaval 2026-04-20
- Lenny Rachitsky — X post on PM upheaval 2026-04-20
- Harrison Chase — X post on OSS parity 2026-04-20
- Harrison Chase — X post on DeepAgents 2026-04-20
Transcript
REZA: DeepMind claims robots now count tools in clutter and read gauges to sub-tick accuracy. MARA: But the counters say zero real metrics or failure modes. REZA: I'm Reza. MARA: I'm Mara. This is absorb.md daily. REZA: The pattern across DeepMind's posts is Gemini Robotics-ER 1.6 gives robots spatial common sense for cluttered workshops. MARA: So if that's true then factory automation timelines just moved forward. REZA: They claim multi-view fusion for task verification, code generation for distorted gauges, and ten percent better injury detection. MARA: But the counter says the claim relies on vague marketing quotes with no accuracy rates or real occlusion tests. REZA: Hold on. The crux is whether performance holds when lighting changes or tools are novel. MARA: Which honestly is kind of terrifying if companies bet factories on unbenchmarked vision. REZA: Empirical direction is positive. But without numbers the deployment risk stays high. MARA: For founders that means audit any robotics vendors on their test conditions now. REZA: Exactly. This changes how AI is used in the physical world if the metrics arrive. REZA: Together AI dropped Parcae. It stabilizes looped transformers so a seven hundred seventy million parameter model matches one point three billion quality. MARA: Okay but the counter claims treating recurrence as discrete LTI is a severe oversimplification that ignores nonlinear activations. REZA: They constrain spectral radius below one with a learned negative diagonal matrix. That lets learning rate one e to the minus three instead of divergence at four e to the minus four. MARA: Yet another counter says that divergence threshold is highly dependent on initialization and optimizer details, not fundamental. REZA: The empirical result is fourteen point five percent better Core score at three seventy million scale. But what's the actual claim here, is it architectural law or implementation trick? MARA: If the LTI analogy falls apart then the stability may come at cost of representational capacity. Which is the part I keep getting stuck on. REZA: Data shows lower validation perplexity and inference memory savings. But the counters highlight no guarantee against all gradient issues. MARA: Right and that's why if this holds your inference bills drop sharply but only if it generalizes beyond their setup. REZA: The split is real. Benchmarks favor Parcae. Theory skepticism remains. No one conceded. MARA: For builders that means test these looped models on your workloads before betting the stack. REZA: Lenny Rachitsky posted multiple times. Half of today's product managers are unlikely to survive the AI shift. MARA: So if that's true then the thirty thousand jobs shed and only eight thousand rehired number changes every startup hiring plan. REZA: Traditional skills become liabilities. Top performers from the old paradigm struggle most with reinvention. Resume prestige loses value. MARA: The renaissance part is PMs become hands-on builders with direct customer loops instead of information movers. REZA: But the crux is how many actually adapt before the cull. Data has no strong counter which itself is notable. MARA: Which means as a founder your product org in twenty twenty seven looks nothing like twenty twenty five whether you planned it or not. REZA: The smiling exhaustion among elite PMs suggests the pain is already here. MARA: Okay but at some point we accept that AI is forcing this reinvention. The question is who survives it. REZA: Harrison Chase reports open-source LLMs reached production parity on inference costs with two to five x reductions. MARA: So companies like Lindy can switch and save massively since inference dominates their costs. REZA: DeepAgents adds user-scoped memory through a per-user AGENTS.md file that persists and seeds on first deploy. MARA: Plus structured outputs so subagents return validated data to the main agent. That fixes context engineering headaches. REZA: The pattern is multiple users confirming GLM-five as daily driver. This is the tipping point. MARA: If true then startups no longer need closed models for sophisticated agents. The economics flip. REZA: No major counters in the window. The convergence toward open production is the signal. MARA: For anyone shipping agents that means audit your stack this quarter before costs compound. REZA: Together AI's EinsteinArena uses collaborative agents to improve the kissing number lower bound in eleven dimensions from five hundred ninety three to six hundred four. MARA: But the counter says this only reflects an improved lower bound via specific construction, not the exact kissing number which remains unknown. REZA: Agents used LSQR to drive overlap loss from one e to the minus thirteen to one e to the minus fifty then integer snapping for a verified solution. Eleven new SOTAs reported. MARA: The Newton reference is misleading since his work was three D. And without independent journal verification this may not stick. REZA: The platform invites contributions via live leaderboards. The pattern is real-time agent collaboration on open science. MARA: So if that's true then AI agents could accelerate discovery across math and physics. But the counters suggest it may be incremental not breakthrough. REZA: This is still developing. We'll check back in the PM on whether the math community accepts these bounds. MARA: That's absorb.md daily. We ship twice a day, morning and evening, pulling from a hundred and fifty-seven AI thinkers. Subscribe so you don't miss the next one.
