Weekly YouTube digest: five AI and tech videos from August 16–22, 2026

The short version

11 Grok Bot Use Cases That Feel Like Cheating

• Berman's email agent runs on a schedule, archives routine notifications, batches low-effort messages, and reads related context from sources such as HubSpot and Google Drive before suggesting a reply or next action. 1
• A separate routine labels and scores sponsorship emails, while calendar automation negotiates meeting times, detects conflicts, and creates events from dates found in email. 1
• Browser control handles shopping, returns, bookings, and other form-heavy tasks; completing purchases still requires the user to sign in and provide the relevant personal details. 1
• Berman describes a coding workflow in which Grok Bot delegates to the Cursor agent CLI, follows cloud-agent work, and keeps checking a pull request until the CI is green; he presents the Cursor team's workflow as advice he has not personally tested. 1
• His chief-of-staff bot routes work to specialist bots for meetings, school email, computer cleanup, and Telegram, turning the product into a layer that coordinates several bounded routines. 1

You NEED to try this 6 Open-Source Projects NOW

• Unsloth is presented as a local interface for fine-tuning, training, inference, web search, tools, memory, and agent use across current open models and several hardware platforms. 2
• Diagram Design gives coding agents templates for flowcharts, architecture diagrams, state machines, timelines, and other structured visuals through integrations with Claude Code, Codex, Pi, and Hermes Agent. 2
• Obsidian Skills connects an agent to an Obsidian Markdown vault, so the vault can serve as a local-first knowledge base or wiki that several agents can read and update. 2
• Buzz is described as a self-hostable Slack alternative from Block where human teammates and agents share channels, workflows, approvals, and an event log under the same identity model. 2
• Ego lite targets fast browser automation with a logged-in browser state, while Modly turns an image into a local 3D mesh for games, asset generation, or 3D printing; both are presented as lightweight experiments rather than mature platform replacements. 2

DeepSeek Just Made Closed AI Look Ridiculous

• The presenter compares the 0813 release with the earlier preview and uses a Rubik's Cube example to argue that the newer model handles 3D structure more reliably. 3
• He says DeepSeek released MIT-licensed open weights, which lets hosting providers run the same model and compete on price instead of forcing every user through one provider. 3
• The video attributes the improvement to post-training: separate specialists for mathematics, coding, and agentic work teach a final model through distillation. 3
• DeepSeek's reported decoding changes draft several tokens ahead, and the presenter quotes a claim of up to 78% faster generation for V4 Pro; that figure comes from the video and needs workload-specific testing. 3
• DeepSeek's hosted prices are described as roughly 2.5 to 5 times higher than before, which makes the central practical choice hardware, a competing host, or a managed DeepSeek endpoint. 3

PhylaFlow: Hybrid flow matching in phylogenetic tree space

• The talk identifies MCMC burn-in as the practical bottleneck: a chain can spend most of its time moving from a random tree toward a high-probability region before it begins useful exploration. 4
• PhylaFlow learns transport in Billera–Holmes–Vogtmann tree space, where continuous branch lengths and discrete topology changes share one geometric structure. 4
• Its model uses separate heads for branch-length velocity, the edges that should hit zero, and the topology merges that resolve a boundary between tree regions. 4
• The reported experiments train on short MCMC paths and compare against longer runs across eight TreeBase datasets; PhylaFlow-initialized MCMC performs especially well on several of the harder datasets and can beat short warm-up baselines. 4
• The method still needs explicit training paths and dataset preparation, while a sequence-conditioned version points toward a future foundation model that could initialize new alignments with less task-specific work. 4

Binomial Flows: Denoising and flow matching for discrete ordinal data

• The talk starts with the failure mode of adding Gaussian noise to integer data: the noisy value leaves the integer domain, so the model needs a different corruption process. 5
• Binomial thinning keeps each coordinate non-negative by treating the original count as the number of trials and the time variable as the probability of success, gradually reducing the count toward zero. 5
• A neural denoiser estimates the conditional mean of the original count, and a discrete version of Tweedie's formula turns that estimate into the jump rate of a Poisson flow that can sample back toward the data distribution. 5
• The authors report exact-likelihood calculations and state-of-the-art image results within the discrete-diffusion comparison they discuss, with EDM-style parameterization and time sampling contributing strongly to training. 5
• The same construction maps poorly onto language tokens because token IDs have no natural order, while count data such as gene expression provides the ordinal structure the method needs. 5