Good morning, and welcome back to Techzip AI Newsletter.

Today’s issue starts with a hard constraint: HBM shortages are already pushing Chinese AI-chip prices higher, while DeepSeek is redesigning V4.1-Flash around lower memory use. Anthropic’s 2030 scenarios widen the lens, asking what happens when model capability meets real workflow redesign, with the rest of the issue tracking agent safety, search, spending and the systems around the models.

DeepSeek has released V4.1-Flash, a 552-billion-parameter model that activates 8 billion parameters while reading input and 16 billion while generating an answer. The company says its context memory needs one-quarter as much HBM and one-eighth as much SSD capacity as the previous generation. The model is live through the API and supports image input.

The interesting part is how it uses memory. Around 197 billion Engram parameters act like a lookup table, while only four layers write compressed context memory for the rest to share. The model supports contexts of up to one million tokens, a clear design choice for tasks that run across many turns.

If the HBM and SSD figures hold up under independent testing, serving costs will depend on more than which model is "smarter." DeepSeek also says tests by multiple parties put V4.1-Flash ahead of V4-Pro on performance, cost, speed and total runtime. Those remain company claims, not independently verified results.

Reuters reports that Chinese AI chipmakers are raising prices as supplies of high-bandwidth memory (HBM) tighten. Huawei is said to be quoting more than 250,000 yuan for its Ascend 950DT, up 20-50% from quotes two months ago; Cambricon and other rivals are also raising prices.

HBM now makes up a large share of an AI accelerator's cost. Since tighter export controls on advanced HBM took effect in December 2024, Chinese chipmakers have relied more heavily on unofficial channels, where memory can cost several times what buyers outside China pay. Huawei's older Ascend 950PR has risen from roughly 60,000 yuan to more than 80,000 yuan, while the Ascend 910C board is now above 110,000 yuan.

This is bigger than a price-list change. Suppliers are reallocating scarce GPUs: Iluvatar CoreX has doubled shipments to ByteDance to 100,000 units and diverted chips originally meant for its own use. Compute capacity is increasingly a question of memory availability; a faster chip roadmap does not help if the HBM cannot arrive with it.

Anthropic's Economics team has published an interactive model for growth, jobs and wages through 2030. Instead of treating AI as an automatic productivity shock, it breaks occupations into bundles of tasks and asks which ones are augmented, automated, left unchanged or newly created. Its three scenarios range from an impact similar to the internet, to AI doing half of knowledge work, to AI doing almost all of it. GDP ends up 1.6%, 8.3% or 32.4% higher than in an economy without AI, depending on the scenario; these are model conditions, not forecasts.

The gains are not evenly shared. In the middle scenario, knowledge-worker wages are roughly flat; in the extreme scenario, they fall by more than 10%, unemployment rises beyond typical recession levels, and labor's share of income drops to 45.2% versus 54.8% for capital. What AI can do is only half the story; how quickly companies and workers redesign their work matters just as much.

This is a scenario model, not a prediction. The authors note that it leaves out policy responses, business cycles, aggregate demand and financial shocks. Its useful question is simpler: how many real workflows have to change before AI capability shows up in the economy?

Here's what's happening with the Groundfloor Preferred Note: while the market moves on headlines, this rate doesn't move once you commit. Lock in a 7% fixed annualized return over a defined 6-month term, paid out monthly — income that isn't tied to earnings surprises, rate cuts, or the next correction. Groundfloor has a 100% repayment track record on Notes since launching them in 2018, and has paid out over $38M in Notes interest alone. New investors get a $100 bonus when they get started.

Invest Now

In a long process, an AI assistant often fails because it does not know which step it is on. A better approach is to map the process as connected steps: what comes first, what allows the work to move forward, and where to return when something fails.

Do not give the assistant the entire map every time. Give it the part around the current step, what just happened, what comes next and which condition allows the transition. This keeps irrelevant instructions out of the way and helps the assistant stay on track.

After each run, compare a successful attempt with a failed one. If the same step keeps breaking, fix the relationship between steps or the transition condition instead of adding another page of instructions. Let the assistant suggest changes, but keep a change only if it also works on a separate test; otherwise, the process will grow more complicated after every fix.

The point is not to create a magically self-improving assistant. It is to turn failures into targeted improvements while keeping the process understandable and controlled.

Jukan argues that adding GPUs or memory will not be enough if substrates, small components and wafer testing remain constrained; meaningful capacity may not arrive until 2028. Read it as a supply-chain view that moves the compute bottleneck beyond the main chip and into the whole production system.

DAIR summarizes Microsoft research on a 4-billion-parameter coding assistant trained through trial and error across roughly 1,500 programming environments, without copying the full behavior of a larger model. The useful idea is to keep generating tasks that are just hard enough for the current system; choosing the right difficulty matters more than producing endless synthetic data.

Omar Sar0 summarizes a paper showing that training a weaker model on an expert model's full working path, inside an optimized tool setup, can reduce performance by 4-30 points across seven tests. The proposed fix is to rewrite only the failed turns instead of forcing the weaker model to copy the stronger model's entire planning style. The practical lesson is simple: swapping the model while keeping the same tools and coordination layer is not automatically an upgrade. The surrounding system can determine how well the assistant actually works.