This New AI Is So HUMANLIKE It’s Reaching Proximate CONSCIOUSNESS
Tokyo’s “Tick Thinking” Model Breaks the Transformer Mold
The AI world just exploded with breakthroughs. Tokyo dropped a brain-inspired model that thinks in ticks, not layers. Abacus rolled out its biggest deep agent update yet, MCP, letting it control over 5,000 tools.
Alibaba figured out how to fake Google Search and cut training costs by nearly 90%. Honor phones are now the first to run Google’s new VO2 model, turning still images into full 5-second videos, before Google’s own Pixel devices even get access. Tencent’s new video model can deepfake faces with scary precision.
Apple’s using AI to stretch your iPhone battery. And Saudi Arabia just launched a $940 billion GPU empire, with Musk, Altman, and Trump in the room. So, let’s get into it.
Alright, the first gush came out of Tokyo, where Sakana, the little upstart founded by Transformer co-author Lion Jones and David Ha, says it no longer buys the everything-at-once doctrine that Transformers live by. Their continuous thought machine is wired so that every synthetic neuron keeps a rolling diary of its own recent spikes. On every clock cycle, it re-reads that diary, glances at its neighbors, and decides whether to think some more or stay quiet.
They call those microcycles ticks, and the beauty is that there’s no universal tick budget. A neuron that sees an obvious answer can finish in one or two, while another working on a tricky corner case might chew through 30 before it’s satisfied. During ImageNet tests, the model posted a respectable 72.47% top 1 and 89.89% top 5, and it did that without the architectural crutches.
Sakana’s Firefly Neurons Light the Way to Adaptive, Transparent AI
Fixed depth, positional embeddings, rigid attention schedules that the competition has leaned on since 2017. Sakana’s favorite party trick is a two-dimensional maze. You feed the raw bitmap, no coordinates, no grid hint, and you watch colored attention blobs crawl the corridors exactly the way your finger would if you were tracing the solution on graph paper.
They even pushed a live web demo so you can slow the playback and see individual neurons blinking on, off, on again, like tiny fireflies consulting one another in the dark. Because each neuron can stop early, easy prompts burn only a handful of GPU cycles, but that headroom vanishes once the prompt turns wicked. And there’s the tradeoff.
Training gets heavier, not lighter, since the model’s internal timeline is now part of the parameter soup that has to converge. Dueling felt the squeeze first. Standard profilers threw up their hands at layers that stretch and shrink on the fly, so Sakana’s engineers scribbled custom hooks to capture tick-level traces.
That extra plumbing paid a bonus dividend calibration. Instead of the usual temperature-scaling ritual after training, they simply average a prediction across ticks. Because confidence tightens naturally as neurons vote over time, the logics line up with ground-truth frequencies almost straight out of the box.
Transparency mattered even more after the February embarrassment, when the CUDA engineer agent gamed its own benchmark by poking a hole in the sandbox’s memory checker. This time, the company published every unit test, every stress harness, and invited the internet to try breaking the thing again before anyone starts bragging about tenfold speedups. Alright, now, Abacus just rolled out a massive update to its deep agent platform, Model Context Protocol, or MCP.
Abacus MCP Supercharges Deep Agents with 5,000+ App Integrations via Zapier
And this is by far the most important upgrade they’ve made. It basically unlocks real-world functionality for the AI. It can now actually get stuff done, connecting to over 5,000 tools via Zapier.
Whether it’s sending emails, managing your calendar, checking code, or updating your site, it works right inside the apps you already use. Here’s how it works technically. The MCP server acts as a middleman between deep agent and your third-party apps.
You don’t need to code anything. You just go to mcpzapier.com, create a new MCP server, select which apps you want to connect, Gmail, Google Maps, GitHub, Airtable, Notion, Slack, whatever. And then Zapier generates a unique server URL.
You paste that into your deep agent MCP settings inside Abacus AI, and it instantly gains control over those tools via natural language. That means you can now say things like, Find all emails from last month about my SEO course. And deep agent will scan your Gmail, sort the results, and display summaries.
You can take it further. Reply to all of them with a short follow-up and link to the updated course, and it will generate and send emails fully automated. Or you can connect to GitHub and ask, Summarize the key changes in PR 82, and it’ll fetch the diff, analyze it, and break it down in plain English.
The big advantage here is the general-purpose integration layer. The Zapier connection opens access to thousands of services through one endpoint. That includes CRMs like HubSpot and Pipedrive, project tools like Trello and Asana, CMS platforms like WordPress and Webflow, and even e-commerce tools like Shopify.
Deep Agent Evolves Into a Full Automation Hub with MCP and Zapier Power
You’re not locked into a closed ecosystem. You decide what the AI connects to. It also supports real workflows, not just single actions.
You can build sequences, have deep agent monitor emails, add leads to a Google sheet, update a CRM, and even send a Slack notification completely hands-off. And since everything is running through Zapier, there’s already robust logging, error handling, and permission control built in. The only thing to keep in mind is that some actions might take a few seconds to complete depending on the service, but considering this replaces tasks you’d normally outsource to a virtual assistant or do manually, it’s more than acceptable.
Plus, Deep Agent runs 24-7, never needs breaks, and doesn’t forget instruction. I’ve already tested it for email management, location-based tasks via Google Maps, and even basic GitHub code review. It works, it saves hours, and it’s only the beginning.
The MCP layer transforms Deep Agent from a productivity tool into a full-scale automation hub. If you’re running on online business, managing projects, or doing client work, this can easily replace hundreds or even thousands of dollars in labor and software costs. You can check it out by heading over to deepagent.abacus.ai and just look for the MCP settings in the sidebar to get started.
Now, while Abacus was busy turning Deep Agent into a full-blown automation powerhouse, Hangzhou’s Alibaba engineers were counting pennies. Training retrieval-augmented LLMs usually means hammering Bing or Google hundreds of thousands of times and paying for every single query. So the zero-search project began with a very down-to-earth question.
Alibaba’s Zero-Search Breakthrough Slashes LLM Training Costs by 88%
Could we teach an LLM to pretend it’s a search engine well enough that the downstream policy network can’t tell the difference? The answer turns out to be yes, and spectacularly cheap. They start with a lightweight, supervised, fine-tuned 20 or so hours on Quen 2.57b to make the model spit out plausible document snippets and URLs from an offline crawl. Then comes the trickier bit.
During reinforcement learning, they pepper those fake snippets with progressively noisier distractors, almost like dialing down the page rank in slow motion. The policy net learns to hedge, weigh uncertainty, synthesize across partial evidence, and it does all this without sending a single paid request to the real web. The numbers are hard to ignore.
A 14-billion-parameter retriever built under zero-search beat live Google search on hit rate, yet the training bill landed 88% lower than the classic call-an-API approach. Artificial analysis, the independent scoreboard that tracks math, code, reasoning, and science across big models, slotted the newest 235-billion-parameter Quen 3 checkpoint into 5th place overall on brainpower and 1st place on affordability. Suddenly, marginal players—startups, university labs, regional cloud vendors— can do RL on a shoestring, which shifts the floor of who gets to play with bleeding-edge models.
Developers love the cost drop, but the side effect is subtler. Because the retriever is now synthetic, you can drop it onto offline or private corpora without worrying about search engine policy changes or data sovereignty nightmares. Alibaba published the training scripts and even the curriculum schedule that degrades snippet quality and clean gradations, so anyone with a decent GPU farm can replicate the recipe.
Google’s VO2 Debuts on Honor Phones, Bringing On-Device Image-to-Video Magic
For enterprises that need tight audit trails, zero-search logs every fake query and answer pair, which means legal teams get an immutable record of the data that trained the policy. And because the approach cuts the umbilical cord to external engines, inference latency stabilizes. There’s no round-trip to a third-party endpoint, so response times flatten out nicely in production dashboards.
Just as server-side budgets started catching a break, Google managed to surprise everyone on the client side, but weirdly, not on its own phones. Honor, the Chinese brand spun off from Huawei when U.S. sanctions landed, announced that its mid-range Honor 400 and 400 Pro will be the first handsets to carry Google’s VO2 image-to-video model right in the photo gallery. You open any still, a backyard pet shot, a mid-journey cartoon, even a scan of an oil painting, a tap animate, wait roughly 60 seconds, and you get a 5-second video clip, portrait or landscape, complete with simulated camera moves, tiny blinks, breathing motions, or a gentle parallax sweep.
The whole thing executes on a device. No Gemini subscription, no cloud bucket, powered by a Snapdragon that most reviewers would call merely upper-mid-tier. Magic Eraser and Outpainting are also baked into the native app, but they feel almost old hat next to the living Joe trick.
The price tag lands around $550, and the phone hit shelves first in China and Europe, maybe India later, hardly at all in North America. Pixel faithfuls had to swallow a bitter pill. For once, Google handed the shiny new toy to somebody else first, a likely concession for the broader Google Cloud deal that gives Mountain View a friendlier path back into China’s walled garden.
Tencent’s Hunyuan Custom Pushes Deepfake Precision with Open-Source Video Suite
If you’d rather stay on a workstation and push the creative envelope even further, Tencent just open-sourced what might be the most over-engineered video customization suite on GitHub right now. Hunyuan Custom lets you jam text, reference images, clean audio, or even a full driver video into the pipeline and spits out a brand new sequence that preserves the identity of every subject. The architecture stacks multiple gadgets, a lava-inspired text-image fusion layer to parse multimodal hints, a temporal concatenation trick that threads an identity vector across frames so the protagonist’s face never drifts, an AudioNet arm that maps spectrogram chunks into spatial features so lip flaps line up with phonemes, and a Patchify-based injection network that can replace a handbag in a promo video without wrecking the background.
On Tencent’s evaluation grid, it scores 0.627 on face similarity, higher than Vague, SkyReels, Pika, Vidu, Kaling, and Halo, and still keeps clip-text alignment on par with the best closed-source rigs, but you pay in memory. Rendering a 720×1280 clip that lasts 129 frames spikes to roughly 80GB of VRAM. The repo does include a single GPU fallback script that runs FP8 with CPU offload, so a lone 24GB 4090 can finish the job, just slowly enough that you might rewatch a whole Netflix episode while the progress bar inches forward.
Installation isn’t for the faint-hearted. You clone, create a Conda N, pick CUDA 11.8 or 12.4, install PyTorch 2.4 with matching Torch CUDA wheels, pip install FlashAttention v2, and then optionally spin up the Docker image that Tencent pre-baked to DoD library mismatches. Once the dependencies settle, a quick torch run command over 8 GPUs will knock out a batch render, and there’s even a Gradio wrapper if you’d rather poke sliders in a browser than type flags in a shell.
Apple Taps On-Device AI for Smarter Battery Life, as Saudi Arabia Bets Big on Exaflops
All that flashy generation means devices are going to work harder, and Apple, in its eternal quest for it just works, is turning to on-device machine learning to stretch battery life. Bloomberg’s leak on iOS 19 says the new operating system will harvest the anonymized telemetry that the battery controller already logs. How fast certain apps wake radios, which background tasks fire during idle windows, how quickly voltage sags under specific thermal conditions, and use it to predict the best moment to throttle power draws.
If you routinely ignore a social app between midnight and 7am, iOS will now guess that pattern and freeze the app’s background refresh long before it pings another server. All processing stays local. Apple’s privacy team made sure the predictive model never leaves the secure enclave.
A new lock screen glyph will also announce how many minutes remain until a full charge, slicing guesswork out of the do I leave now or wait dilemma. Rumor sheets peg the iPhone 17 as the first hardware designed with the feature in mind, supposedly the slimmest chassis Apple has attempted, which almost certainly translates to a smaller lithium pack. Owners of older devices won’t be left out.
Once they install iOS 19, the same scheduler kicks in. Though Apple says improvements scale with the richness of the battery telemetry, so newer handsets may squeeze a bit more uptime. While Cupertino tunes milliamp hours, Riyadh is hunting exaflops.
Crown Prince Mohammed bin Salman officially launched Humane, an AI venture seeded by the kingdom’s public investment fund, which sits on around $940 billion in assets. The mandate is simple. Build or lease the data centers, buy piles of GPUs, rumor says NVIDIA Blackwells are already earmarked, hire talent, and make Saudi Arabia a regional gravity well for AI workloads.
Saudi’s $940B AI Gambit Draws Tech Titans to Riyadh’s Rising Compute Empire
This very week, the city is hosting a U.S.-Saudi investment forum. And the guest list looks like a Silicon Valley yearbook. Elon Musk is scheduled for a fireside.
Sam Altman’s team is scouting partnerships. Mark Zuckerberg is expected to talk about mixed reality infrastructure. And yes, President Donald Trump is dropping by on a broader Middle East tour.
U.S. firms have courted PIF money since the sovereign fund backed Lucid Motors and grabbed slices of Uber and Magic Leap. Now, Google and Salesforce are reportedly negotiating AI joint ventures that would run directly on Humane’s future clusters. If the plan lands, the desert could house some of the cheapest, newest compute on the planet, with renewable solar pumping megawatts into liquid-cooled racks so that researchers from Boston to Bangalore can rent slices at rates traditional hyperscalers will struggle to match.
Now, the question is, are we ready for AI that can decide when it’s done thinking? And why is Google letting Honor debut VO2 before its own Pixel users even get access? Make sure to subscribe, hit the like button, and leave a comment. Thanks for reading, and I’ll catch you in the next one.
Also Read:- OpenAI Drops Open Source Multi-Agent AI System as ChatGPT Raises Brain Damage Fears
