Experience
Education
Other similar profiles
Explore more posts
-
Alex Sarlin
Andrew Goldman, EVP, HMH Labs | 2026 prediction: “The focus of AI in education will shift from outputs to outcomes.” In this clip from our year-end predictions series, Andrew shares a hopeful but important prediction about where AI needs to go next. AI adoption is accelerating because it genuinely helps educators who are stretched across too many demands. It solves real, immediate pain points. But Andrew argues that the next phase can’t stop at productivity or outputs alone. In 2026, the real question will be how AI supports learning outcomes in meaningful, measurable ways, not just whether it makes tasks faster or easier. Listen to more predictions from Andrew and other EdTech leaders in Year-End Special Part 1: Reviewing 2025 + 2026 AI & EdTech Predictions 👉 https://lnkd.in/e2eZUxGM #EdTech #AIinEducation #LearningOutcomes #FutureOfLearning
2 Comments -
Jesse Landry
San Francisco’s startup grid just got a new surge, and the voltage is high. Plural Energy, once known as Calica, closed a $7.13M oversubscribed seed round led by Paradigm, with Maven 11, Volt Capital, and Neoclassic Capital on the ticket. That’s not just capital, it’s validation. When Coinbase’s co-founder and a former Sequoia partner line up, it signals this isn’t a pitch deck dream. It’s execution with teeth. Founded in Dec 2022 by Adam Silver, Alexander Fong, and Jason G., with Kent Kolze stepping in as CTO in 2024, the team reads less like a garage and more like a command center. Silver comes from Deloitte and ServiceNow. Fong’s fingerprints stretch across IBM, Sonder, and Fast Break Labs. Grissino drives business development with pace. Kolze, fresh from Citadel, now channels trading precision into tokenizing renewables. Add Howard Steinberg as CLO after stints at Goldman Sachs, Scotiabank, Forge Global, and LTSE, and you’ve got a regulatory shield few can match. The proof is in numbers. $300M in pre-screened #solar and #battery projects ready. The team has managed over $4B in renewable portfolios, covering 3.5 GW of assets. Their debut Ace Portfolio exceeded targets, hit the market in Nov 2024, and delivered blockchain-based distributions that turned heads. Forty thousand investors are already on the waitlist for the first #onchain launch. That’s not hype, that’s demand. Plural Energy’s edge is ruthless efficiency. Blockchain and #smartcontracts aren’t buzzwords, they cut two percent off project sponsors’ cost of capital by removing intermediaries and automating compliance. For billion-dollar projects, that two percent isn’t pocket change, it’s survival. As a registered transfer agent and now broker-dealer through the acquisition of NorthStar Financial Services Group, LLC (rebranded Plural Brokerage), the company owns the vertical. Infrastructure meets innovation without skipping steps. The road ahead is wired for scale. Plural Energy plans to #tokenize $100M+ in clean energy assets in 2025 while lowering investment minimums so access isn’t gated by institutions. Long term, assets flow into DeFi protocols as collateral and find liquidity on #decentralizedexchanges. #Cleanenergyfinance becomes programmable, global, and accessible. Plural isn’t singular in ambition. It’s the intersection of tech, capital markets, and renewable energy. Congrats to Adam Silver, Alexander Fong, Jason Grissino, and Kent Kolze for locking in this seed round and setting the stage for what’s next. Credit to Paradigm, Maven11, Volt Capital, and Neoclassic Capital for recognizing that tokenized energy isn’t a trend, it’s the next asset class. #Startups #StartupFunding #EarlyStage #VentureCapital #SeedRound #Blockchain #DeFi #FinTech #Energy #EnergyTech #CleanEnergy #Compliance #Infrastructure #Technology #Innovation #TechEcosystem #StartupEcosystem If software engineering peace of mind is what you crave, Vention is your zen.
3 Comments -
Eric Seufert
New paper from Meta's content moderation team introducing industrial-scale RL for policy-violation classification. Scaled content moderation is a thorny problem: policies are nuanced and generally fluid, data is sparse and expert labels are expensive to produce, and the threat vector is constantly evolving. Meta's content moderation team explored the use of RL to fine-tune policy classification models and found that it outperformed SFT "on tasks requiring complex, policy-grounded reasoning" while achieving up to two orders of magnitude higher data efficiency under constrained labeling regimes. The team also found that the RL approach exhibited sigmoid-like scaling efficiencies as more compute and data were utilized. The paper establishes content moderation as a classification problem: given content C and policies p_1...p_N, the content is prohibited if y_i=1 for any p_i. The team built an RL algorithm using Group Relative Policy Optimization (GRPO), an alternative to Proximal Policy Optimization (PPO) that eschews value functions and uses relative advantages across sampled responses and updates the policy using groupwise comparison. This achieves higher stability in light of data sparsity, which is an acute problem in content moderation. The team uses Monte Carlo sampling to produce a smoother probability distribution of violation scores, since the initial approach produced a bimodal distribution. The team also introduces prompt adjustments that force the model to revisit previous decisions in the reasoning stage. What the team found was that RL underperforms when classification is merely a function of verifiable accuracy, as it is with eg., math problems or code production. In soft-policy regimes like content moderation, reasoning plays an outsized role in determining rewards, and so reasoning completeness becomes the principal source of model viability. The expanded reasoning process increases the probability of selecting a correct response, improves consistency across rollouts, and reduces variance in outcomes. This was verified using best-case correctness (pass@N) and consistency across samples (maj@N). The result is that the RL-only post-trained model outperformed the SFT post-trained model in low-data environments, using 1/10th as many samples. And the team found that increased token usage, rollouts, and compute led to consistent performance gains, although it tapered off at some saturation point. Paper linked below.
1 Comment
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content